Monte Carlo Methods in Finance Peter J¨ackel 24th November 2001
ii
Contents Preface
ix
Mathematical Notation
xiii...
42 downloads
1134 Views
15MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
Monte Carlo Methods in Finance Peter J¨ackel 24th November 2001
ii
Contents Preface
ix
Mathematical Notation
xiii
1
Introduction
1
2
The mathematics behind Monte Carlo methods
5
2.1
A few basic terms in probability and statistics . . . . . . . . . . . . . . . . . . . . . .
5
2.2
Monte Carlo simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
3
2.2.1
Monte Carlo supremacy . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
2.2.2
Multidimensional integration . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
2.3
Some common distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
2.4
Kolmogorov’s strong law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
2.5
The central limit theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
2.6
The continuous mapping theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
2.7
Error estimation for Monte Carlo methods . . . . . . . . . . . . . . . . . . . . . . .
19
2.8
The Feynman-Kac theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
2.9
The Moore-Penrose pseudo-inverse . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
Stochastic dynamics
23
3.1
Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
3.2
Itˆo’s lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
3.3
Normal processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
3.4
Lognormal processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
3.5
The Markovian Wiener process embedding dimension . . . . . . . . . . . . . . . . .
27
3.6
Bessel processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
3.7
Constant elasticity of variance processes . . . . . . . . . . . . . . . . . . . . . . . .
28
3.8
Displaced diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
iii
iv 4
5
6
7
CONTENTS Process driven sampling
31
4.1
Strong versus weak convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
4.2
Numerical solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
32
4.2.1
The Euler scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
32
4.2.2
The Milstein scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
4.2.3
Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
4.2.4
Predictor-Corrector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
4.3
Spurious paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
4.4
Strong convergence for Euler and Milstein . . . . . . . . . . . . . . . . . . . . . . .
37
Correlation and co-movement
41
5.1
Measures for co-dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
5.2
Copulæ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
5.2.1
The Gaussian copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
5.2.2
The t-copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
5.2.3
Archimedean copulæ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
50
Salvaging a linear correlation matrix
57
6.1
Hypersphere decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
6.2
Spectral decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
6.3
Angular decomposition of lower triangular form . . . . . . . . . . . . . . . . . . . .
61
6.4
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
62
6.5
Angular coordinates on a hypersphere of unit radius . . . . . . . . . . . . . . . . . .
63
Pseudo-random numbers
65
7.1
Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
7.2
The mid-square method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
70
7.3
Congruential generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
71
7.4
Ran0 to Ran3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
72
7.5
The Mersenne twister . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
73
7.6
Which one to use? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
74
v
CONTENTS 8
9
Low-discrepancy numbers
75
8.1
Discrepancy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
76
8.2
Halton numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
77
8.3
Sobol’ numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
78
8.3.1
Primitive polynomials modulo two . . . . . . . . . . . . . . . . . . . . . . . .
79
8.3.2
The construction of Sobol’ numbers . . . . . . . . . . . . . . . . . . . . . . .
80
8.3.3
The Gray code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
82
8.3.4
The initialisation of Sobol’ numbers . . . . . . . . . . . . . . . . . . . . . . .
83
8.4
Niederreiter (1988) numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
86
8.5
Pairwise projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
86
8.6
Empirical discrepancies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
88
8.7
The number of iterations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
93
8.8
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
93
8.8.1
Explicit formula for the L2 -norm discrepancy on the unit hypercube . . . . . .
93
8.8.2
Expected L2 -norm discrepancy of truly random numbers . . . . . . . . . . . .
94
Non-uniform variates
95
9.1
Inversion of the cumulative probability function . . . . . . . . . . . . . . . . . . . .
95
9.2
Using a sampler density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
96
9.3
9.4
9.2.1
Importance sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
99
9.2.2
Rejection sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
Normal variates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 9.3.1
The Box-Muller method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
9.3.2
The Neave effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Simulating multi-variate copula draws . . . . . . . . . . . . . . . . . . . . . . . . . 104
10 Variance reduction techniques
107
10.1 Antithetic sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 10.2 Variate recycling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 10.3 Control variates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 10.4 Stratified sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 10.5 Importance sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 10.6 Moment matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 10.7 Latin hypercube sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
vi
CONTENTS 10.8 Path construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 10.8.1 Incremental . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 10.8.2 Spectral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 10.8.3 The Brownian bridge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 10.8.4 A comparison of path construction methods . . . . . . . . . . . . . . . . . . . 124 10.8.5 Multivariate path construction . . . . . . . . . . . . . . . . . . . . . . . . . . 126 10.9 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 10.9.1 Eigenvalues and eigenvectors of a discrete-time covariance matrix . . . . . . . 129 10.9.2 The conditional distribution of the Brownian bridge . . . . . . . . . . . . . . . 132
11 Greeks
135
11.1 Importance of Greeks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 11.2 An Up-Out-Call option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 11.3 Finite differencing with path recycling . . . . . . . . . . . . . . . . . . . . . . . . . 137 11.4 Finite differencing with importance sampling . . . . . . . . . . . . . . . . . . . . . . 139 11.5 Pathwise differentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 11.6 The likelihood ratio method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 11.7 Comparative figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 11.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 11.9 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 11.9.1 The likelihood ratio formula for Vega . . . . . . . . . . . . . . . . . . . . . . 147 11.9.2 The likelihood ratio formula for Rho . . . . . . . . . . . . . . . . . . . . . . . 149 12 Monte Carlo in the BGM/J framework
151
12.1 The Brace-Gatarek-Musiela/Jamshidian market model . . . . . . . . . . . . . . . . . 151 12.2 Factorisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 12.3 Bermudan swaptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 12.4 Calibration to European swaptions . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 12.5 The Predictor-Corrector scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 12.6 Heuristics of the exercise boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 12.7 Exercise boundary parametrisation . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 12.8 The algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 12.9 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 12.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
CONTENTS 13 Non-recombining trees
vii 175
13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 13.2 Evolving the forward rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177 13.3 Optimal simplex alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 13.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 13.5 Convergence performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 13.6 Variance matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184 13.7 Exact martingale conditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 13.8 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 13.9 A simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 13.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 14 Miscellanea
193
14.1 Interpolation of the term structure of implied volatility . . . . . . . . . . . . . . . . . 193 14.2 Watch your cpu usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 14.3 Numerical overflow and underflow . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 14.4 A single number or a convergence diagram? . . . . . . . . . . . . . . . . . . . . . . 198 14.5 Embedded path creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 14.6 How slow is exp()? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 14.7 Parallel computing and multi-threading . . . . . . . . . . . . . . . . . . . . . . . . . 201 Bibliography
205
Index
215
viii
CONTENTS
Preface This book is about Monte Carlo methods and close relatives thereof. It is about the application of traditional and state-of-the-art sampling techniques to problems encountered in the world of modern finance. The approach I take is to explain methods alongside actual examples that I encountered in my professional working day. This is why there may be a bias towards applications to investment banking and derivatives pricing in particular. However, many of the methods presented here equally apply to similar mathematical problems that arise in market risk management, credit risk assessment, the insurance businesses, strategic management consultancy, and other areas where the effect of many unknown variables (in the sense that we can only make assumptions about their probability distributions) is to be evaluated. The purpose of this book is to be an introduction to the basic Monte Carlo techniques used nowadays by expert practitioners in the field. There are so many areas of Monte Carlo methods in finance that any attempt to try and provide a book on the subject that is both introductory and comprehensive would have meant many years of (part-time) writing. Instead, in order to fill the need for an introductory text more timely, I decided to rather focus on the issues most pressing to any novice to financial Monte Carlo simulations and to omit many of the more advanced topics. The subjects not covered include the whole family of Markov Chain techniques, and almost all of the recent advances in Monte Carlo methods tailored specifically for the pricing of American, Bermudan, or any other derivative contract whose ideal value is given by the maximal (discounted) expected payoff over all possible exercise strategy, i.e. by finding the truly optimal exercise strategy. An exception to this is perhaps the identification of a suitable exercise boundary optimisation for the purpose of Bermudan swaption pricing in the Brace-Gatarek-Musiela/Jamshidian framework presented in chapter 12. At the same time, though, I have tried to include most of the presently used techniques that enable the practitioner to create rather powerful Monte Carlo simulation applications indeed. Whilst I always endeavour to explain the basic principles of the particular problem to which a technique is applied, this book is not meant to be an introduction to financial mathematics. I assume that the reader either has background knowledge in the relevant areas, or could follow up the given references for a deeper understanding of the financial and/or economical reasoning behind specific mathematical assumptions. After all, this is not a book about the reasoning behind option pricing. This is a book about mathematical and numerical techniques that may be used for the solution of the mathematical equations that were derived by experts in financial theory and economics. I do not attempt to give a justification for the assumption of complete markets, market efficiency, specific stochastic differential ix
x
Preface
equations, etc.; I leave this up to the authors of the excellent books on those issues subject in the literature [Hul97, Mer90, Reb98, Wil98]. Instead I have focussed on the implementational aspects of Monte Carlo methods. Any Monte Carlo method will invariable have to run on a computing device, and this means that numerical issues can be of paramount importance. In order for this book to be of some practical value to the practitioner having to implement Monte Carlo methods, I made the attempt to link the fundamental concepts of any one technique directly to the algorithm that has to be programmed, and often explicitly in terms of the C++ language, often taking into account aspects of numerical analysis such as round-off error propagation etc.
The nature of the subject of this book is strongly entwined with the concept of convergence. In general, Monte Carlo methods give us at best a statistical error estimate. This is in contrast to various other numerical methods. A Monte Carlo calculation is typically of the following structure: carry out the same procedure many times, take into account all of the individual results, and summarise them into an overall approximation to the problem in question. For most Monte Carlo methods (in particular those providing serial decorrelation of the individual results), we can choose any subset of the individual results and summarise them to obtain an estimate. The numerically exact solution will be approached by the method only as we iterate the procedure more and more times, eventually converging at infinity. Clearly, we are not just interested in a method to converge to the correct answer after an infinite amount of calculation time, but rather we wish to have a good approximation quickly. Therefore, once we are satisfied that a particular Monte Carlo method works in the limit, we are naturally interested in its convergence behaviour, or, more specifically, its convergence speed. A good part of this book is dedicated to various techniques and tricks to improve the convergence speed of Monte Carlo methods and their relatives. In order to present the reader not just with a description of the algorithms, but also to foster an intuitive grasp of the potential benefit from the implementation of a specific technique, we have attempted to include many diagrams of typical convergence behaviour: frequently these are used to highlight the differences between the performances of different methods. In particular where such comparisons are made, we often display the convergence behaviour as a function of cpu time used by the different methods since the human user’s utility is much more closely related to the time elapsed until a calculation of sufficient accuracy has been completed rather than to the number of actual iterations carried out.
You may wonder why there is no explicit chapter on option pricing, considering that that’s one of the most immediate applications of Monte Carlo methods in finance. As it happens, there isn’t one chapter on option pricing, but every chapter is written with option pricing in the mind. My foremost use of Monte Carlo methods has been in the area of derivatives pricing. Since a lot of the examples I give are directly with respect to option valuation, I considered it unnecessary to have a chapter on the subject by itself, only to repeat what is written in other chapters already. I hope the reader will agree with me.
Preface
xi
Acknowledgements The first person the author wishes to thank is undoubtedly Dr. Henrik Neuhaus. It was his vision and inspiration that brought together the Group Risk Division of Nikko Securities Global Holdings in Victoria Street, London, opposite New Scotland Yard. During the little over a year of its existence, prior to the equity buy-in of Citigroup into Nikko (Japan) and the subsequent demise of the London operation of Nikko Europe, I had the pleasure to learn a great deal from Henrik and to be given guidance as I was finding my financial legs and beginning to employ Monte Carlo techniques for derivatives modelling and derivatives model validation. Equally, my thanks go to Yasuo Kusuda, who was always ready to share his business experience, financial expertise, and mathematical knowledge with me. At the same time, I met Prof. Carol Alexander whose ubiquitous enthusiasm to share with other people the results of her own research, and whose continuous encouragement for others to do the same, gave part of the motivation to start this project. I would like to thank her for repeated invitations to speak at her courses on market risk which allowed me to gain a feeling to what extent there may actually be use for a book specifically on Monte Carlo techniques. And of course, I am deeply indebted to Dr. Dongning Qu and Dr. Russ Bubley who I had the pleasure to work with in my days at Nikko. I had many interesting discussions with both of them on the subject of derivatives pricing and Monte Carlo methods. Another person I thank is Dr. Riccardo Rebonato from whom I learned about the BGM/J market model approach to interest rates derivatives. Dr. Rebonato founded what was first the NatWest Group Risk Quantitative Research Centre, then the Royal Bank of Scotland Group Risk Quantitative Research Centre, and appointed my former two colleagues Dr. Chris Hunter and Dr. Mark Joshi to whom I also express my thanks. Chapter 12 is largely based on my experience gained in our time of developing BGM/J models. I would also like to thank my colleagues in the Financial Engineering team of Commerzbank Securities, London. It has been most enjoyable to work with them, and I had many ideas for this book as a result of discussions with my colleagues there. The following fellow financial engineers of mine requested to be mentioned by name, and it is my pleasure to do so: ‘CurryMan’ Dr. Peter Robinson, Dr. Rhys Mainwaring, and Dr. Carl Seymour. Furthermore, I am indebted to Dr. Alexander McNeil for his advice on some of the finer details of copula theory. Also, my thanks go to Mike Curran for his invaluable patience in reading the first draft version of this book with such exhaustive attention to the detail. And finally, there are a number of people who I would like to thank for their support in this project or guidance they may have given me in the past. The complete list would be rather long, so, for the sake of brevity, I name but a few: My D. Phil. supervisor Prof. Tom Mullin, my friend Dr. Paul Wilmott, and my grandmother.
xii
Preface
Disclaimer In this book, I have tried to give an introductory overview of Monte Carlo methods in finance known to expert practitioners, and, in places, I may not always have given due credit to all the pioneers who contributed to this borderline area of mathematics and finance. Wherever I fail to give explicit reference to the original inventor of any given method, this is not to mean I wish to pretend that it is my own development, it is merely my own laxness about the whole issue of referencing and citations. In fact, I may use phrases like ‘I present below’ etc. repeatedly, but they just stand for their literal meaning, namely that I present, not that I claim having invented the particular method. I did consider it much more important to focus on an as-good-as-possible explanation of the techniques and mathematics, rather than spending the time on acribic research through a whole string of references to establish who actually was the originator of the subject at hand. I include a rather-too-long list of references at the end of the book, and I did try to reference and cite wherever I could see a direct link, but I may have failed many great researchers in the field of Monte Carlo methods by not referencing them in the right places, or not referencing them at all. Mea culpa, mea maxima culpa.
Mathematical Notation int[x]
the largest integer that is equal to or less than x.
A∧B
A or B are true, or both.
m XOR n
bitwise exlusive OR operation on the binary representation of the two integers m and n, also known as binary addition without carry.
m ⊕2 n
the same as XOR.
∝
is proportional to.
f −1 (x)
the inverse of the function f (x), i.e. the unique solution for y in f (y) = x, if it exists.
∼
has distribution, i.e. meets the righ-hand-side probability density law.
Z
the set of integers including 0.
N
the set of natural numbers, i.e. the positive integers 1, 2, . . .
R
the set of real numbers.
:=
is defined as.
≡
is identically equal to.
≈
is approximately.
x>
transpose of the vector x.
∂x f
∂f . ∂x
2 ∂xy f
∂2f . ∂x∂y
xiii
xiv
Mathematical Notation
P [A]
the probability of some event, A, occurring.
Eψ [X]
the expectation of the random variable X with distribution density ψ.
hXi
the same as E[X].
Vψ [X]
the variance of the random variable X with distribution density ψ.
Covψ [X, Y ]
the covariance of the random variables X and Y under the joint distribution density ψ(X, Y ).
Corrψ [X, Y ]
the correlation of the random variables X and Y under the joint distribution density ψ(X, Y ). E[|X| p ].
p p
||X||p
the Lp -norm of X defined as ||X||p :=
U(a, b)
the uniform distribution with equal probability density
1 b−a
everywhere on
the open interval (a, b). N (µ, σ)
the cumulative normal distribution with mean µ and standard deviation σ.
ϕ(z)
the standard normal distribution density function, i.e. ϕ(z) =
N(z)
the standard cumulative normal distribution probability function, Rz
i.e. N (z) =
1 2 √1 e− 2 z 2π
ϕ(z 0 ) dz 0 .
z 0 = −∞
N−1 (p)
the inverse cumulative normal probability function. i.e. N(z) = p
=⇒
N−1 (p) = z. δij
the Kronecker symbol which is 1 if i = j, otherwise 0.
δ(x − x0 )
the Dirac density whose singularity is located at x0 .
h(x − x0 )
the Heaviside function whose discontinuity is located at x0 . The connection to the Dirac density is h(x − x0 ) =
Rx
δ(x0 − x0 ) dx0 .
x0 =−∞
1{expression} √ B S, K, σ T
1 if expression is true, otherwise 0. the Black call option formula.
xv
Mathematical Notation √ √ √ ln ( S/K ) ln ( S/K ) 1 1 B S, K, σ T = S· N σ√T + 2 σ T − K· N σ√T − 2 σ T .
σBlack (v, S, K, T )
the Black call option implied volatility, i.e. √ B S, K, σ T = v =⇒ σBlack (v, S, K, T ) = σ
σ(t)
some kind of instantaneous volatility.
σ ˆ (T )
an implied volatility given by the root mean square of its instantaneous counterpart, i.e. σ ˆ (T ) =
B(p, q)
R1 u=0
σ 2 (u · T ) du.
the Beta function, also known as Euler’s integral of the first kind, is defined as B(p, q) :=
R1 0
tp−1 (1 − t)q−1 dt. Rx
tp−1 (1 − t)q−1 dt.
Bx (p, q)
the incomplete Beta function, defined as Bx (p, q) :=
Γ(x)
the Gamma function, also known as Euler’s integral of the second kind, is defined as Γ(x) :=
R∞ 0
0
e−t tx−1 dt.
Its relation to the Beta function is B(p, q) =
Γ(p)Γ(q) . Γ(p+q)
Note that Γ(n + 1) = n! for n ∈ N. Γ(x, y) Ce
R∞ the incomplete Gamma function, is defined as Γ(x, y) := y e−t tx−1 dt. n P 1 Euler’s constant is defined as Ce := lim − ln n and is approxk n→∞
k=1
imately Ce ≈ 0.577215664901532. ζ(s)
n P
1 s. n→∞ k=1 k
Euler’s Zeta function defined as ζ(s) := lim
ented as the following infinite product ζ(s) ≡
Q
It can also be repres-
1 1−p−s
the values of all prime numbers. ∂( f (x) ) ∂( x )
the Jacobian matrix of the transformation x → f (x).
|A|
the determinant of the matrix A.
where p takes on
xvi
Mathematical Notation
Chapter 1 Introduction We are on the verge of a new era of financial computing. With the arrival of ever faster desktop computers, clustering technology, and the tools to utilize spare cpu cycles from a variety of sources, computer processing power has expanded dramatically. This expansion, coupled with the development of new numerical methods is making techniques which were previously considered to be prohibitively computationally expensive not only feasible but the method of choice. There are countless examples of mathematical problems faced by the quantitative analyst which used to require employing either analytical approximations or numerical truncations. This usually meant that the mathematical model that was ideally desired could not be used given the constraints of having to integrate its implementation into existing systems which would not tolerate long calculation times for individual models. Even where integration into a corporate IT infrastructure was not required, the user of a model might have limited patience or business needs that necessitate a comparatively speedy response. Whilst a front office trader in a securities and derivatives house would usually not be concerned if a model built for him by his quantitative support team were to take anything between a few seconds to possibly several minutes in order to price a complex structure with calibration to market data, having to wait several hours would make the model unviable. This is the reason that to date, when the choice between a conceptually superior but numerically less tractable model and one that lends itself to easy analytical evaluation has to be made, very often, the easier model is chosen even though everyone agrees that it oversimplifies the matter at hand. With the aid of ever faster computers and improved Monte Carlo techniques, however, we are nowadays more often than not in the position to use the conceptually superior and more realistic modelling approach. Even where we appear to have analytical solutions it is often desirable to have an alternative implementation that is supposed to give the same answer. The reason for this is that very often the final formulæ of the analytical solution, although mathematically of very elegant appeal, prove to be numerically difficult to evaluate since they still involve one or two one-dimensional integrations, contour integrals in the complex plane, back transformations from Laplace- or Fourier space, or simply the evaluation of special functions that are numerically very difficult to produce, such as the confluent hypergeometric functions or the modified Bessel function of the second kind for large parameters. 1
2
Introduction
Examples of the above include option pricing for generic distributions [CM99], stochastic volatility models [Hes93], analytical approximations for Asian options [FMW98], Variance Gamma processes [MS90, MM91, MCC98], or the Constant Elasticity of Variance process. And finally, there are myriads of situations when we are very satisfied with a certain approximation but would like to have a Monte Carlo simulation tool for comparison, just in case. A good example for this is the difference between continuously monitored and discretely monitored payoff functions such as barrier or American style digital options. There is an excellent approximation by Broadie, Glassermann, and Kou [BGK99] that relates the price V of all simple continuous barrier contracts in a Black-Scholes framework of constant volatility σ to their discretely monitored equivalent where the period of time between monitoring points is τ as follows. Say the barrier level as it is specified on a term sheet of a discretely monitored contract is Hd and we have an analytical formula1 Vc = F (Hc ) for the continuously monitored equivalent contract in a Black-Scholes framework setting. Then, the value of the discretely monitored contract is approximately ζ( 1 ) √ ± √2 σ τ
Vd ≈ F (e
2π
· Hd ) .
(1.1)
with ζ( 1 ) √ 2 = 0.5825971579390106702051771641876311547290938701987... 2π
(1.2)
In other words, the discretely monitored contract is approximated as a continuously monitored contract with a shifted barrier level given by ζ( 1 ) √ ± √2 σ τ
Hc ≈ e
2π
· Hd .
(1.3)
The sign in the exponent in equations (1.1) and (1.3) is selected according to whether the initial spot level is above or below the threshold barrier. Barriers that are above the initial spot level need to be upwards amended when we go from discrete to continuous monitoring, and so the positive sign is used. For barriers that need to be crossed from above for an event to be triggered, the negative sign applies. This approximation works extremely well as long as τ is significantly smaller than the remaining time to maturity of the contract and as long as the current spot level is not too close to the barrier, and most exotic derivatives traders are happy to use it under those circumstances. When a given contract comes close to its expiry, though, or the spot level gets dangerously close to a barrier, traders frequently wish to know to what extent would a different, not-so-approximative valuation differ. In a situation like this, a fast-convergence Monte Carlo simulation that also provides the hedge parameters, can make all the difference for the exotic derivative trader. The concept of random sampling as a computational means has long been established. A wellknown example is the idea to calculate the circle number π by randomly placing points in a unit square with an enclosed circle around the centre of that square. In the limit of many samples, the ratio of 1
There are many good sources for exotic option formulæ. One of my favourites is [Hau97].
3 points that happen to be inside the circle to the total number of samples approaches
π . 4
A way to
compute π is thus to carry out this experiment for an ever increasing number of samples, take the hit ratio, and multiply it by 4. An alternative method attributed to the French naturalist Comte de Buffon [Hol59, CK85, Sch74, Ree01] is to draw parallel lines on a board at a distance l between each adjacent pair of them, and to throw pins of length l randomly onto this board. The ratio of all the pins crossing a line to all those that were thrown should converge to Rπ 2 cos α dα 2 0 = Rπ 2 π dα
(1.4)
0
for large numbers of pins. Before the invention of fast computing machines, however, these approaches were rather time-consuming and cumbersome. The history of Monte Carlo methods as a computational method for the calculation of expectations on potentially high-dimensional domains starts in the mid-1940s with the arrival of the first programmable computers. The main originators are considered to be the American mathematicians John von Neumann, Stanislav Ulam, and Nicholas Metropolis [UvN47, MU49, Sob94]. The first published mentioning of the name “Monte Carlo Method” is an article by Metropolis and Ulam in 1949 [MU49]. In this article, the authors explain how they view the calculation of multi-dimensional integrals resulting from the analysis of the transport problem of radiation and energetic particles governed by the Fokker-Planck equation as a statistical problem. Instead of attempting to directly carry out the highdimensional integrations involving the transition probabilities of many possible intermediate events and states by the use of lattice methods, they sampled single chains of events. The name of the method is only mentioned casually: The idea of using a statistical approach at which we hinted in the preceding examples is sometimes referred to as the Monte Carlo method.
In fact, the term had been used before amongst those scientist in allusion to the principality of Monaco that is so famous for its casinos and to the fact that the roulette wheel represented the archetypical random number generator. Another reason why this allusion is rather appropriate is that some of the early mathematicians who contributed greatly to the development of statistics and probability theory did so in the pursuit of gaining riches at the gambling tables. In the same paper, the authors also establish the result that the Monte Carlo method enables us to evaluate the expectations of functionals of certain variables without the knowledge of the distribution of the variables themselves: all that is needed is a description of the process that generates those variables, and off we go! What’s more, in 1949, Metropolis and Ulam already point out that the Monte Carlo method is easily amenable to parallelised computing. The only problem with that was, as one of my colleagues once put it, that few people had more than one machine the size of a tennis court2 called “a computer” readily at their disposal at the time. 2
This is a joke, I don’t actually know what size the computer(s) used by von Neumann, Metropolis, and Ulam were. It is probably safe to assume, though, that they were substantially larger than today’s desktop PCs, and they certainly were a lot harder to come by.
4
Introduction In 1947, a Monte Carlo simulation involved many problems, not only the sheer size of the machine.
Apart from the fact that the mathematical relations between the partial differential equations describing the problems the scientists were investigating, the associated multi-dimensional integrals, and their formulation as a stochastic process that can be readily simulated were only just being established, the actual simulation was a rather formidable enterprise itself. In those days, a computer was a rather large machine indeed, and to operate it involved several machine operators. The most advanced input/output device available was a punch card reader and a card puncher. The generation of random numbers was a huge problem and many different ways were explored. It may seem somewhat unreal when we think about it now, but one of the approaches taken at the time was as follows: A special project codenamed RAND was set up [Bro48, Bro51] whose aim, amongst others, was to accumulate a set of numbers that were considered to be sufficiently random for Monte Carlo simulations. RAND devised an electronic roulette wheel by combining a random frequency pulse source with a constant frequency pulse (≈1 Hz) which provided about 100,000 pulses per second. The regularised pulses were passed to a fiveplace binary counter which effectively resulted in the equivalent of a roulette wheel with 32 positions. After a binary to decimal conversion, these numbers were then fed to an IBM punch, thus generating cards of random digits. Exhaustive tests on the so generated set of numbers revealed some slight bias and they were later modified by shuffling and other further processing until they finally met the applied statistical tests [Bro51]. However, let me just give you a feeling for how serious the whole business was with the following quote taken from George Brown’s report [Bro51] on the RAND project published in the National Bureau of Standard’s booklet [Mon51] on the Monte Carlo Method in 1951. Production of random numbers really began on April 29, 1947, and by May 21 there were half a million digits generated. [...] By July 7 there were a million digits available [...]. [...] At this point we had our original million digits, 20,000 IBM cards with 50 digits to a card, [...].
Fortunately, computers, Monte Carlo methods, and algorithms for number generation have come a long way since, and their application to the mathematical problems encountered in modern finance is what this book is about.
Chapter 2 The mathematics behind Monte Carlo methods The theoretical understanding of Monte Carlo methods draws on various branches of mathematics. In this chapter, a few useful facts and results from the areas of probability, statistics, calculus, and linear algebra are summarised. Some of them are going to come in handy in the following chapters, and the remainder is presented here for reference. Most readers will already be familiar with the contents of this chapter, and may as well skip it...
2.1
A few basic terms in probability and statistics
A random experiment is a process or action subject to non-deterministic uncertainty. We call the outcome of a random experiment a draw or a variate from a distribution. An example for this would the flipping of a coin, or to be precise the face shown once the coin has come to rest. A distribution density is a generalised function that assigns likelihood or probability density to all possible results of a random experiment. A synonym for distribution density is probability density function. For our purposes, a generalised function can be an ordinary function, or a linear combination of an ordinary function and any finite number of Dirac densities δ(x − x0 ). The Dirac density is the equivalent of assigning a finite probability to a single number on a continuous interval. This means, the Dirac density δ(x − x0 ) is zero everywhere where it is defined, and strictly speaking undefined at x0 . However, its Rx integral is given by the Heaviside function, i.e. h(x − x0 ) = δ(x0 − x0 ) dx0 , which is zero for x < x0 x0 =−∞
and one for x > x0 . The relationship between the probability that the outcome X of a random experiment is an element of some set S and the distribution density ψ(x) of the random experiment is Z Pr [x ∈ S] =
ψ(x) dx . S
5
(2.1)
6
The mathematics behind Monte Carlo methods We call the set of all attainable outcomes X of a random experiment the domain D(X) of the
random experiment. Whenever D(X) is an ordered set, i.e. when we can decide whether any one element is less than any of the other elements of D(X), we define the cumulative probability function or just cumulative for short as Zx Ψ(x) =
ψ(x0 ) dx0 = Pr [X < x] .
(2.2)
x0 =inf(D)
All distribution densities are normalised, i.e. Z ψ(x) dx = 1 .
(2.3)
D
Note that in the probability sciences the cumulative is also referred to plainly as the distribution function or simply the distribution. Since this can lead to situations of ambiguity, I always explicitly state whether I am talking about the distribution density or the cumulative. The expected value of a quantity subject to uncertainty is the probability weighted average. Our notation for the expected value of a quantity f with respect to a probability density ψ is Eψ [f ]
:
expected value of f with respect to ψ.
(2.4)
Very often, there is no ambiguity about the relevant distribution. In this case we may just write E[f ]. Alternatively, there may be a parametric description of the distribution of f . This means that f is actually a function of some other uncertain variable x. Given the distribution density of x, say ψ(x), we would denote the expectation of f as Eψ(x) [f (x)]. This is just to say Z ∞ f (x)ψ(x) dx . Eψ(x) [f (x)] =
(2.5)
−∞
Analogously, we will need to make statements about the variance of a given uncertain quantity f — intuitively this is the variablity of a quantity from its expected value. The variance of f is denoted by V[f ]. It is defined by V[f ] = E (f − E[f ])2 = E f 2 − (E[f ])2
.
(2.6)
The standard deviation of a random variate or distribution is defined as the square root of its variance. The covariance of two quantities f and g is defined as Cov[f, g] = E[f · g] − E[f ] · E[g]
(2.7)
and based on covariance is the concept of linear correlation Cov[f, g] . Corr[f, g] = p V[f ] V[g]
(2.8)
The correlation coefficient is by construction always in the interval [−1, 1]. The n-th moment of a distribution is defined as Z n Eψ(x) [x ] = xn ψ(x) dx .
(2.9)
7
2.2 Monte Carlo simulations Jensen’s inequality For any convex function g(x) and any distribution density ψ(x) the following inequality holds: Eψ [g(X)] ≥ g (Eψ [X])
(2.10)
p E[XY ] ≤ kXkp · kY k p−1
(2.11)
H¨older’s inequality For any two random variates X and Y ,
for any p > 1. Minkowski’s inequality For any two random variates X and Y , kX + Y kp ≤ kXkp + kY kp
(2.12)
for any p ≥ 1.
2.2
Monte Carlo simulations
The most common use of Monte Carlo simulations in finance is when we need to calculate an expected value of a function f (x) given a specified distribution density ψ(x) over x ∈ Rn , Z v = Eψ(x) [f (x)] = f (x)ψ(x) dxn .
(2.13)
In a strictly mathematical sense, Monte Carlo maximisation (i.e. the search for the maximum value of f (x) for x in a given domain D) too, can be formulated as in equation (2.13). This can be seen as follows. Provided that f is non-negative everywhere in D, define sZ ms = s [f (x)]s ψ(x) dxn .
(2.14)
D
The maximum of f in D is then given by lims→∞ ms . Whilst in practice one would not use this limit procedure1 to determine the actual maximum of a function f in a domain D, it is a very useful formal trick when the objective is to derive a partial differential equation for the value of a derivative contract that depends on the maximum value of an underlying asset over a given time interval [Wil98]. The easiest form of Monte Carlo integration of integrals like equation (2.13) can be summarised as follows. 1
One would, of course, simply keep track of the maximum value of f seen thus far as we iterate through more and more samples.
8
The mathematics behind Monte Carlo methods · Establish a procedure of drawing variates x from the target distribution density ψ(x). · Set up a running sum variable double RuningSum=0;, a running average variable double RunningAverage;, and a counter variable unsigned long i=0;.
· Draw a variate vector xi and evaluate fi := f (xi ). · Add the computed function value to RunningSum. · Increment i, i.e. ++i;. · Set the running average to RunningAverage = RunningSum/i;. This gives us the Monte Carlo estimator vˆN :=
N 1 X f (xi ) . N i=1
(2.15)
· Keep iterating until either the required number of iterations has been carried out, or a specific error estimate (see section 2.7) has decreased below a predetermined threshold. For more sophisticated methods when a procedure to draw variates exactly from the target distribution ψ(x) is either not available or computationally too expensive, see chapter 9.
2.2.1
Monte Carlo supremacy
The concept of Monte Carlo integration becomes ever more appealing as the dimensionality of the domain increases. How this can be becomes clear as we look at the convergence behaviour of lattice methods in comparison to Monte Carlo methods. For a given d-dimensional hypercube of volume λd the error εLattice of approximating the integral of a given function over the hypercube’s interior by linear approximation decreases like O(λ2 ). In a regular grid over a d dimensional domain D, the number of points N relates to the subdivision length λ like N ∝ ( 1/λ)d
(2.16)
which implies λ∝N
−1/ d
.
(2.17)
Thus, for a fixed number N of evaluations of the function which is to be integrated on a d-dimensional domain, the relative error of a lattice method relates to N and d according to: εLattice (N, d) ∝ N −
2/ d
.
(2.18)
Picking the sample points at random, however, the relative error relates to N and d according to εMC (N, d) ∝ N −
1/ 2
(2.19)
9
2.3 Some common distributions
and is thus independent of the dimensionality. Note that equations (2.18) and (2.19) imply that in the limit of many sampling points, the Monte Carlo method breaks even with lattice techniques as of d = 4, and outperforms regular grid methods for d > 4. For much higher dimensionalities, lattice methods become utterly unusable whence the term curse of dimensionality.
2.2.2
Multidimensional integration
This brings us to the question: what is the total mathematical dimensionality of a Monte Carlo simulation? Well, for simple problems such as the expected return from a portfolio given a multi-dimensional joint distribution for the returns from all of the underlying assets out to a specific time horizon, it would just be the number of those underlying assets. However, if we are carrying out a Monte Carlo evaluation of an expectation of a function(al) of discretised paths of a set of financial assets over some monitoring dates, then the state vector x represents the realisations of all involved assets at all future times of relevance, and the density ψ(x) is the joint distribution density of all those realisations. The dimensionality of the integration domain is then d=k·l .
(2.20)
with k being the number of assets and l being the number of time horizons in the time discretisation of the evolution paths.
2.3 Some common distributions In general, most of the problems that we deal with when we resort to Monte Carlo methods are based on a quantity whose distribution is more or less well known analytically from the mathematical formulation of the chosen model. In this section, some of the more frequently used distributions are listed. Uniform A random experiment that can have a continuum of outcomes X which are all equally likely or equiprobable is said to have a uniform distribution. Most uniform distributions that we encounter in finance are either on a single closed interval [a, b] or a single open interval (a, b). In the former case, we say X ∼ U[a, b]
which means
D(X) = [a, b] ,
X ∼ U(a, b) which means
D(X) = [a, b] .
and in the latter
In either case, the probability density is given by 1 ψ(x) = · 1{x∈D} . b−a
10
The mathematics behind Monte Carlo methods
Normal Very often, the quantity of interest depends on underlying variates that are normally distributed with mean µ and standard deviation σ. In this case, we say that a variate X has the distribution N (µ, σ), i.e. X ∼ N (µ, σ). The distribution density function is 2 1 (x−µ) 1 ψ(x; µ, σ) = √ e− 2 σ2 . σ 2π
(2.21)
Of particular common usage are standard normal variates. A variate Z is said to be a standard normal variate if Z ∼ N (0, 1) .
(2.22)
I usually denote the standard normal distribution density function by 1 2 1 ϕ(x) = √ e− 2 x . 2π
(2.23)
The standard normal distribution is symmetric around the origin and thus all odd moments vanish. For even moments, we have k Y 2k 2k (2k − 1)! 1 E X = √ Γ(k + /2) = (2m − 1) = (k−1) 2 (k − 1)! π m=1
The standard cumulative normal probability function is Z x N(x) = ϕ(x0 ) dx0 .
for
k>0.
(2.24)
(2.25)
−∞
The function N(x) relates to the error function via 1 z N(z) = 1 + erf √ 2 2
(2.26)
and √ erf(x) = 2N( 2x) − 1 .
(2.27)
At the time of this writing, most compiler systems provide a highly accurate and fast implementation of the error function. These implementations are usually based on rational Chebyshev interpolation schemes for the error function which have been around for a long time, are well understood, highly reliable, and very accurate [Cod69]. One such implementation is included on the accompanying CD, just in case your compiler system and libraries lack it. Since the cumulative normal distribution function is very frequently used in financial calculations, it is advisable to implement it in terms of a highly accurate and efficient error function by virtue of equation (2.26). Equally, for the inverse cumulative normal function z = N−1 (p), there are several numerical implementations providing different degrees of accuracy and efficiency [MZM94]. A very fast and accurate
11
2.3 Some common distributions
approximation is the one given by Boris Moro in [Mor95]. The most accurate whilst still highly efficient implementation currently freely available, however, is probably the algorithm by Peter Acklam [Ack00]. When allowing for an additional evaluation of a machine-accurate cumulative normal distribution function, Acklam’s procedure is able to produce the inverse cumulative normal function to full machine accuracy by applying a second stage refinement using Halley’s method [ST95, Pic88]. A C++ implementation of Acklam’s method is also included on the CD. Bernoulli or dichotomic A variate X is said to be dichotomic or Bernoulli distributed if it can take on only two discrete values, say A and B. Since the associated probabilities must add up to 1, we have p := Pr [A] = 1 − Pr [B] . Binomial The binomial distribution is generated by repeating a Bernoulli experiment, say, n times. The Binomial variate is then the number of times that one of the two states was attained. The probability that A occurs k times is in n repetitions of a Bernoulli experiment is n k Pr [X = k] = p (1 − p)n−k . k Thus, the distribution density of the Binomial variate X is n X n k ψ(x) = p (1 − p)n−k δ(x − k) . k k=0
(2.28)
(2.29)
The cumulative probability function results by integration to the same sum as in (2.29), only with the Dirac densities replaced by Heaviside functions. Geometric The geometric distribution describes the number of times one has to repeat a Bernoulli experiment until state B occurs for the first time. Since Pr [X = k] = pk−1 (1 − p) ,
(2.30)
we have ψ(x) =
∞ X
pk−1 (1 − p)δ(x − k)
(2.31)
k=1
for the geometric distribution, and similarly with Heaviside instead of Dirac densities for the cumulative.
12
The mathematics behind Monte Carlo methods
Poisson This distribution is closely related to the concept of point processes. Assume that there is a probability intensity λ that in any one time interval dt, a certain event happens. In other words, the probability of the said event occurring in any time interval dt is dp = λdt. The probability that this event has occurred k times over a time span [0, T ] is given by (λT )k . k!
Pr [X = k] = e−λT
(2.32)
The density and cumulative result again by summing up probability-weighted Dirac densities and Heaviside, respectively. Exponential Assume again that there is a probability of dp = λdt of an event occurring over any one time interval. The probability density of the time one has to wait until the next event happens is given by the exponential distribution density ψ(t) = λe−λt .
(2.33)
Ψ(t) = 1 − e−λt
(2.34)
The cumulative probability is
and the moments of the exponential distribution are k! . λk
(2.35)
Γ(α + β) α−1 xα−1 (1 − x)β−1 x (1 − x)β−1 = Γ(α)Γ(β) B(α, β)
(2.36)
Etk [=] Beta Beta variates are distributed according to ψ(x) =
for x ∈ [0, 1] and some α, β > 0. It can be seen as a generalisation of the uniform distribution in the sense that it is a smooth curve on the interval [0, 1], identically 0 at either end, and by a suitable choice of the parameters α and β it can be made very similar to the uniform distribution. Expectation, variance, and moments of the beta distribution are E[X] =
α , α+β
V[X] =
αβ , (α + β + 1)(α + β)2
Γ(α + k)Γ(α + β) and E X k = . (2.37) Γ(α + β + k)Γ(α)
The cumulative probability function of the beta distribution is Ψ(x) =
Bx (α, β) . B(α, β)
(2.38)
13
2.3 Some common distributions Gamma The density function of the gamma distribution reads ψ(x) =
xα−1 − βx e Γ(α)β α
(2.39)
for x ≥ 0 and some α, β > 0. For α = n and β = λ1 , the waiting time until n events have occurred in a Poisson process is gamma distributed according to (2.39). Expectation and variance of the gamma distribution are E[X] = αβ
(2.40)
V[X] = αβ 2 .
(2.41)
Its cumulative is Ψ(x) = 1 −
Γ (α, x/β ) . Γ(α)
(2.42)
Chi-square Define X as the sum of the squares of ν independently drawn standard normal variates, i.e. X :=
ν X
zi2
with
zi ∼ N (0, 1) .
(2.43)
k=1 2
The variate X follows the law of the χ distribution density with ν degrees of freedom ν
x x 2 −1 ψ(x) = ν ν e− 2 Γ( 2 )2 2
(2.44)
and has the cumulative probability function Ψ(x) = 1 −
Γ ( ν/2, x/2) . Γ( ν/2)
(2.45)
In general, ν doesn’t need to be an integer. The expectation and variance are E[X] = ν
(2.46)
V[X] = 2ν .
(2.47)
The χ2 distribution is a special case of the gamma distribution with α = ν/2 and β = 2. Student’s t Take two independent variates Y and Z. The first one, Y is a standard normal variate Y ∼ N (0, 1), and the second one is drawn from a χ2 distribution with ν degrees of freedom, i.e. Z ∼ χ2ν . Then, the quotient Y X := q
(2.48) Z ν
14
The mathematics behind Monte Carlo methods
satisfies Student’s t distribution ψ(x) =
Γ( ν+1 ) √ 2 ν νπ Γ( 2 )
− ν+1 2 x2 2 − ν+1 1 + 2 ν x 1+ = √ . ν ν B( ν/2, 1/2)
(2.49)
All the moments up to order ν exist. Since the distribution is symmetric around 0, its expectation is 0. For ν > 2, the variance and higher order even moments for k < ν are ν ν−2 k Γ( ν+1 )Γ( ν−k ) 2 2 √ E X = . ν π Γ( 2 ) V[X] =
(2.50) (2.51)
For x ≤ 0, the cumulative probability function is Ψ(x) =
ν B ν+x ( ν/2, 1/2) 2
2B( ν/2, 1/2)
(2.52)
and for x > 0 it is given by virtue of its symmetry as Ψ(x) = 1 − Ψ(−x). Cauchy The Cauchy (also known as Lorentz) distribution is defined by the density function 1 1 π 1 + x2
(2.53)
1 1 atan(x) + . π 2
(2.54)
ψ(x) = and its cumulative probability function is Ψ(x) =
The Cauchy distribution is symmetric around zero and thus all of its odd moments are zero2 . None of the even moments exist. Lognormal Define X := eσz
(2.55)
with z being a standard normal variate z ∼ N (0, 1). The variate X follows the lognormal distribution 1 ln2 x
e− 2 σ2 √ ψ(x) = σx 2π ln x Ψ(x) = N . σ
(2.56) (2.57)
2 This is assuming that the moments are defined by Riemann integration. This is important since the half-sided integrals R ±∞ 2k+1 x ψ(x)dx for some k ∈ N0 diverge for the Cauchy distribution and thus Lebesgue integration would not result in x=0 finite moments of odd order.
15
2.3 Some common distributions The expectation, variance, and moments are 1
2
E[X] = e 2 σ 2 σ2 σ V[X] = e e − 1 k 1 2 2 E X = e2k σ
(2.58) (2.59) (2.60)
The lognormal distribution is probably the most important function in computational finance since it is the solution to the stochastic differential equation describing geometric Brownian motion, which in turn is almost always the first easy choice when the evolution of non-negative quantities is to be modelled. Generalised beta 2 The generalised beta distribution of the second kind [BM87, CDM90], or GB2 for short, is given by a four-parameter density function ψ(x) =
|a|xap−1 a p+q bap B(p, q) 1 + xb
(2.61)
for a, b, p, q, x > 0. The cumulative of the GB2 distribution can be written as Ψ(x) =
B−( x )a (p, 1 − p − q) b
(−1) p B(p, q)
.
(2.62)
Note that (−1)
−c
B−y (c, d) = (−1)
−c−1
Z
−y c−1
t
d−1
(1 − t)
Z dt =
0
y
sc−1 (1 + s)d−1 ds
(2.63)
0
for c > 0 and d < (1 − c) is a real positive number and can be evalutated very efficiently, e.g., by the aid of an adaptive Gauss-Lobatto method [GG00b, GG00a, Gan92]. The moments of the GB2 distribution are B p + ka , q − E X k = bk B(p, q)
k a
.
(2.64)
One of the very useful features of the GB2 distribution is that it allows for a large variety of shapes that are nearly lognormal which is desirable when density functions are used for the extrapolation of volatility smiles of traded options. Pareto The Pareto distribution is skewed and heavy-tailed and frequently used to model rare events. Its density function is ψ(x) =
a xa+1
(2.65)
16
The mathematics behind Monte Carlo methods
for x ≥ 1 and a > 0 and its cumulative probability function is Ψ(x) = 1 −
1 . xa
(2.66)
All moments of order ν with ν < a exist and are given by E Xk =
a . a−k
(2.67)
aba xa+1
(2.68)
Generalised Pareto As a generalisation of (2.65) we have ψ(x) =
for x ≥ b and a > 0 and its cumulative probability function is a b Ψ(x) = 1 − . x
(2.69)
abk E Xk = a−k
(2.70)
The moments are
for k < a.
Weibull This distribution density is given by ψ(x) = γ
xγ−1 − xβγ e β
(2.71)
for x ≥ 0 and γ, β > 0 and its cumulative probability function is xγ
Ψ(x) = 1 − e− β
(2.72)
k E X k = β /γ Γ(1 + k/γ ) .
(2.73)
The moments are
The Weibull distribution is sometimes also referred to as the Frechet distribution.
17
2.3 Some common distributions Gumbel The Gumbel probability density is ψ(x) = abe−(be
−ax +ax
)
(2.74)
for a, b > 0 and x ∈ R. Its cumulative is −ax
Ψ(x) = e−be
.
(2.75)
The first three moments of the Gumbel distribution are 1 (ln b + Ce ) a 2 1 π2 2 2 E X = 2 ln b + 2Ce ln b + Ce + a 6 3 1 π2 π2 3 2 2 3 E X = 3 ln b + 3Ce ln b + 3Ce + ln b + Ce + Ce + 2ζ(3) . a 2 2 E[X] =
(2.76) (2.77) (2.78)
The number ζ(3) ≈ 1.2020569 is also known as Ap´ery’s constant.
Generalised lambda The generalised lambda distribution is also known as the asymmetric lambda distribution. It exists in two forms of parametrisation: the original form by Ramberg and Schmeiser [RS74, RDTM79] and a later rewritten formulation [FMKL88, KDM96]. Unlike most other distributions, it is directly defined in terms of the inverse of the cumulative. In the original parametrisation [RS74], this reads 1 uλ3 − (1 − u)λ4 λ2
Ψ−1 (u) = λ1 +
(2.79)
for u ∈ [0, 1]. Its density is parametrically given by x = Ψ−1 (u) and −1 −1 d −1 λ3 λ3 −1 λ4 λ4 −1 ψ(x) = Ψ (u) = u + (1 − u) . du λ2 λ2
(2.80)
The generalised lambda distribution is very popular for the fitting of data due to its tremendeous flexibility. With the right choice of parameters it can be used to approximate almost all of the previously mentioned distributions. The moments of the generalised lambda distribution can be calculated directly as E Xk =
Z 0
1
k Ψ−1 (u) du.
(2.81)
18
The mathematics behind Monte Carlo methods
The first three moments are 1 1 1 E[X] = λ1 λ2 + − λ2 λ3 + 1 λ4 + 1 2 1 1 2λ1 λ2 2λ1 λ2 1 2 2 E X = 2 λ1 λ2 + + + − − 2 B(λ3 + 1, λ4 + 1) λ2 2λ3 + 1 2λ4 + 1 λ3 + 1 λ4 + 1 1 3λ21 λ22 3λ1 λ2 1 1 3λ21 λ22 3λ1 λ2 E X 3 = 3 λ31 λ32 + + + − − + λ2 1 + λ3 2λ3 + 1 3λ3 + 1 3λ4 + 1 λ4 + 1 2λ4 + 1
(2.82) (2.83)
− 3 B(2λ3 + 1, λ4 + 1) − 6λ1 λ2 B(λ3 + 1, λ4 + 1) + 3 B(λ3 + 1, 2λ4 + 1) . (2.84)
2.4 Kolmogorov’s strong law This fundamental result by Andrey Nikolaevich Kolmogorov was established in the 1920s and is the main mathematical justification for the use of Monte Carlo simulations. Convergence almost surely Let X1 , X2 , X3 , . . . be a sequence of random variates, such as, for instance, the running average of a Monte Carlo simulation. If, given some ξ, for all ε, η > 0 there exists an n0 such that Pr [|Xn − ξ| > ε, ∀ n > n0 ] < η ,
(2.85)
then we say that the sequence {Xn } converges almost surely to ξ which is denoted as a.s.
Xn −→ ξ .
(2.86)
Kolmogorov’s strong law of large numbers Given a sequence of iid, i.e. independent identically distributed, variates ξi with expectation E[ξi ] = µ ,
(2.87)
define their running sum and average as Sn :=
n X
ξi
(2.88)
1 Sn . n
(2.89)
i=1
Xn := Then, a.s.
Xn −→ µ .
(2.90)
19
2.5 The central limit theorem
2.5
The central limit theorem
The central limit theorem is one of the cornerstones of statistics and probability theory. It was first formulated by Laplace. Since then, many great mathematicians such as Chebyshev, Markov, Lyapunov, Khinchin, Lindeberg, and L´evy have given extensions and variations of it. For the sake of brevity, only a rather simplified form is outlined below. Given a sequence of independent identically distributed variates ξi with expectation and variance E[ξi ] = µ
(2.91)
V[ξi ] = σ,
(2.92)
define the running sum Sn as in (2.88). Then, for increasing n, the composite variate Xn :=
Sn − nµ √ σ n
(2.93)
converges in distribution to the standard normal distribution. We denote this as i.d.
Xn −→ N (0, 1) .
(2.94)
A practical example for this is the approximation of a Gaussian variate by summing up 12 uniform (0, 1) variates. Since the variance of uniform variates amounts to 1/12, the denominator in (2.93) equals unity and the approximation is simply the sum of twelve uniforms minus 6. Note that this is not a highly accurate approximation and should only be used to establish ballpark estimates. A diagram is given in figure 2.1 for comparison3 .
2.6 The continuous mapping theorem Given a sequence (Xn , Yn ) that converges in distribution to (X, Y ), i.e. i.d.
(Xn , Yn ) −→ (X, Y ) ,
(2.95)
and let φ(x, y) be a continuous map. Then i.d.
φ(Xn , Yn ) −→ φ(X, Y ) .
(2.96)
2.7 Error estimation for Monte Carlo methods Given a Monte Carlo estimator vˆ as the average of many individual draws of the random variate V , i.e. N 1 X vˆN = vi , N i=1 3
My thanks to Dr. Carl Seymour for carrying out the calculations for this diagram.
(2.97)
20
The mathematics behind Monte Carlo methods 0.4
0.35
0.006 standard Gaussian [sum of 12 uniforms (0,1)] - 6 difference
0.005
0.004 0.3 0.003 0.25
0.2
0.001
difference
density
0.002
0 0.15 -0.001 0.1 -0.002 0.05 -0.003
0
-0.004 -4
-2
0
2
4
Figure 2.1: Approximating a standard normal variate by taking twelve U(0, 1) variates and subtracting 6. The uniform number generator used for this diagram is the Mersenne Twister (see section 7.5). we know that each individual evaluation of the estimator itself, for large N , behaves approximately like a normal variate by virtue of the central limit theorem. Assuming that the variance of V is σ 2 , this means i.d.
vˆN −→ N
σ µ, √ N
.
(2.98)
Since vˆN approaches a normal distribution, a statistical measure for the uncertainty in any one simulation result of vˆN is given by the standard deviation of vˆN , namely p σ V[ˆ vN ] = √ . N
(2.99)
In general, we don’t actually know the variance σ 2 of the random variate V whose expectation we are trying to estimate. However, by virtue of the combination of the central limit and the continuous mapping theorem, we can use the variance of the simulation instead as an estimate for σ 2 : v ! !2 u N N u 1 X X 1 σ ˆN = t v2 − vi N i=1 i N i=1
(2.100)
This leads us to the definition of the standard error: σ ˆN N = √ N
(2.101)
21
2.8 The Feynman-Kac theorem
Whenever the standard error is used as an error measure, it is important to remember that it is only of statistical nature. Any single simulation can yield a result well outside the standard error. As a matter of fact, since the total probability mass within one standard deviation either side of the centre of the standard normal distribution is only around 68.3%, approximately every third simulation based on random numbers will have a result outside the standard error margin around the correct solution! In addition, the standard error itself is subject to a statistical error margin. It can be shown, however, that the standard deviation of the standard error scales like
2 σ ˆN /N
and is thus much less significant than the
statistical nature of the standard error as a measure for expected accuracy.
2.8
The Feynman-Kac theorem
This theorem by R. Feynman and M. Kac [Fey48, Kac51] connects the solutions of a specific class of partial differential equations to an expectation which establishes the mathematical link between the PDE formulation of the diffusion problems we encounter in finance, and Monte Carlo simulations. Given the set of stochastic processes dXi = bi dt +
n X
aij dWj
for
i = 1..n ,
(2.102)
j=1
with formal solution Z Xi (T ) = Xi (t) +
T
Z
T
bi dt + t
t
n X
aij dWj ,
(2.103)
j=1
any function V (t, X) with boundary conditions V (T, X) = f (X)
(2.104)
that satisfies the partial differential equation n n X ∂V 1X ∂2V ∂V +g+ cij + bi = kV ∂t 2 i,j=1 ∂Xi ∂Xj ∂X i i=1
(2.105)
with cij :=
n X
aik ajk
(2.106)
k=1
can be represented as the expectation Z T R R − tT kdu − ts kdu V (t, X) = E f (X T )e + ge ds .
(2.107)
t
Hereby, all of the coefficients aij , bi , k, and g can be functions both of time t and the state vector X(t). As with most mathematical theorems, there is a whole host of additional conditions for good behaviour of all the coeffients and functions involved and the reader is referred to, e.g., Karatzas and Shreve [KS91] (page 366).
22
The mathematics behind Monte Carlo methods
2.9
The Moore-Penrose pseudo-inverse
The Moore-Penrose pseudo-inverse [GK65, Alb72, PTVF92] of a matrix A ∈ Rm×n is a very robust way of attempting to solve linear systems that for a variety of reasons may be singular or near-singular. It is based on the method of singular value decomposition of A given by
=
λ1 ·
A
U
λ2 ... ..
. λn
·
.
V>
(2.108)
The entries λi of the diagonal matrix Λ in the centre of the right hand side are called the singular values, and are guaranteed to be positive. The matrices U and V are each columnwise orthonormal. Now, define the diagonal matrix Θ of the same dimensions as Λ by setting its diagonal entries to 1 for λi > λi (2.109) θi := 0 for λ ≤ i for some suitable round-off threshold (which can reasonably safely also be set to exactly zero). We thus obtain the Moore-Penrose pseudo-inverse: > A−1 MP := V · Θ · U
(2.110)
Using the Moore-Penrose pseudo-inverse to solve a linear system Ax = b will result in a least-square fit whenever the system is overdetermined and thus strictly speaking has no solution. For underdetermined systems, it will find the one solution that has no projection onto the nullspace of A. In other words, it will find the solution of least norm and thus avoid the accidental amplification or contribution of noise. The Moore-Penrose pseudo-inverse is the method of choice whenever our objective is to find a vector (or matrix) x that best fits Ax = b
(2.111)
for some vector (or matrix) b with the least possible L2 norm for x. In most situations when we encounter a singular or ill-conditioned system of this nature, this is precisely what suits us best. An excellent reference for further details and the numerical method to compute the singular value decomposition of any matrix is [PTVF92].
Chapter 3 Stochastic dynamics In all aspects of life, we face having to take into account the unknown future. In the world of finance we very often want to quantify the uncertainty of the future, or at least the financial impact of the universe’s possible evolutions we may incur. The mathematical way to describe that a quantity could evolve in various ways with associated probabilities as time passes by is to say that it is subject to stochastic dynamics. Great mathematicians have worked in this area and today we owe tremendeously to Gauss, Wiener, Brown, Levy, Itˆo, and many others for path-breaking discoveries on stochastic processes and stochastic calculus. In this chapter, I briefly summarise some of the more important ones the quantitative analyst commonly encounters. This list is by far not comprehensive, it is merely a little reminder of the whole zoology of stochastic processes that has been studied by mathematician.
3.1
Brownian motion
Brownian motion is ubiquitous in quantitative finance. Wherever we model the evolution of something into the future, the simplest assumption is to say that over any one time step the outcome only depends on the present value of the quantity (Markov property) and that all time steps are equal, i.e. that the structure of the evolution process is the same at all times into the future (stationarity). If in addition to that we demand that the quantity must evolve continuously (without jumps), we necessarily will have to use a Brownian process as the underlying stochastic driver of the quantity we wish to model. The reason for this is that the above set of properties already defines Brownian motion which makes it a beautifully simple concept. This is best summarised in the following theorem taken from [Har90]. Theorem 3.1.1 If Y is a continuous process with stationary independent increments, then Y is a Brownian motion. Harrison continues: This beautiful theorem shows that Brownian motion can actually be defined by stationary independent increments and path continuity alone, with normality following as a consequence of
23
24
Stochastic dynamics these assumptions. This may do more than any other characterization to explain the significance of Brownian motion for probabilistic modeling.
A standard Brownian motion is often also referred to as a standard Wiener process owing to the fact that it was N. Wiener in the 1920s who started the mathematical analysis of Brownian motion. A Wiener process is formally often represented by the differential increment dW or dWt (indicating that this is the increment of the Wiener process happening at time t) and this notation is adopted throughout this chapter. A standard Wiener process has the following properties:1. The expectation of the differential increment dWt at any one point in time t is zero: E[dWt ] = 0
(3.1)
2. The variance of the differential increment dWt at any one point in time t is equal to the associated differential increment in time: V[dWt ] = E dWt2 − (E[dWt ])2 = E dWt2 = dt
(3.2)
This means that the sum of increments, i.e. the value of the process variable Wt , is normally √ distributed with zero mean and a standard deviation of t: Z t √ Wt = dWs ∼ N (0, t) (3.3) s=0
In the design of Monte Carlo algorithms we make frequently use of a more general formulation of the second property of a Wiener process: Z
s
T
f (t) dWt
Z
∼
N 0,
t=0
T
f (t)2 dt
(3.4)
t=0
This results immediately from the fact that variance of a Wiener process is linearly additive. Equation (3.4) is of fundamental importance in many applications since it enables us to bypass the numerical solution of the stochastic integral on the left hand side and directly draw variates that represent the distribution of the entire expression.
3.2
Itˆo’s lemma
One of the most important mathematical tricks from stochastic calculus that we use for the purpose of transforming stochastic differential equations is Itˆo’s lemma. In crude terms, it is the direct consequence of the variance of a Wiener process increment over any infinitesimal time step given by equation (3.2). Itˆo’s lemma states that the total differential of any given function f (X) depending on a stochastic variable X ∈ Rn subject to the system of stochastic differential equations dXi = bi (t, X) dt +
n X i=1
aik (t, X) dWk t
(3.5)
25
3.3 Normal processes is given by df =
n X
n 1 X 2 (∂Xi f )bi + (∂X f )aik ρkl ajl i Xj 2 i=1 i,j,k,l=1
! dt +
n X
(∂Xi f )aik dWk ,
(3.6)
i,k=1
where ρkl =
E[dWk dWl ] dt
(3.7)
is the instantaneous correlation of dWk t and dWl t .
3.3 Normal processes Normal processes are of paramount importance in financial applications. This is not only due to the fact that they are generally very amenable to analytical evaluations, but also that they very often are considered to be a very good starting point for a realistic modelling of the problem at hand. An example is the celebrated Vasicek model of a mean-reverting evolution of the short interest rate towards a longterm equilibrium distribution [Vas77]. This model can be formulated by the stochastic differential equation dr = a(θ − r)dt + σ dW
(3.8)
with a, θ, and σ being constants. Given an initial value for r, i.e. r0 = r(0), the distribution of r(t) can be explicitly calculated. In order to do so, we utilise the so-called method of variation of constants. The first step is to solve the homogeneous differential equation drh = a(θ − rh )dt
(3.9)
rh (t) = θ + c · e−at
(3.10)
to obtain
for some constant c. We now assume that the solution of equation (3.8) is of the same form as (3.10), only that c is not a constant. In other words, we use the Ansatz r = θ + c · e−at
(3.11)
dr = d θ + c · e−at = a(θ − r)dt + e−at dc
(3.12)
and then compare
with the original equation (3.8) which gives us dc = eat σ dW
(3.13)
26
Stochastic dynamics
and thus −at
r(t) = θ + (r0 − θ)e
−at
+e
Z σ
t
eas dWs .
(3.14)
s=0
By virtue of equation (3.4), we can immediately write down the distributional behaviour of r(t) as r r(t) = θ + (r0 − θ)e−at + σ
1−e−2at
z
(3.15)
2a
with z being a standard normal variate or, equivalently, r r(t)
∼
θ + (r0 − θ)e−at , σ
N
1−e−2at
! .
(3.16)
2a
For very small values of the mean reversion speed a, the Taylor expansion of the standard deviation √ is σ t. Thus, initially, the short rate r appears to diffuse like a standard Wiener process, albeit with a drift. For large values of t, however, the distribution converges to a stationary form of N θ, √σ2a . This highlights that mean reverting processes of the type of equation (3.8) (also known as OrnsteinUhlenbeck processes) converge to an equilibrium state for the distribution over long time scales. Equation (3.8) can be extended to allow for a, θ, and σ to be a function of t. The distribution resulting from a so generalised Ornstein-Uhlenbeck process is still Gaussian and can be derived analytically as outlined above, i.e. df = a(t) [θ(t) − f ] dt + σ(t) dWt
(3.17)
leads to −A(t)
f (t) = e
Z f0 +
t A(s)
a(s)θ(s)e 0
with A(s) :=
−A(t)
t
Z
σ(s)eA(s) dWs
ds + e
(3.18)
0
Rs
a(u)du and the distribution reads sZ Z t t σ 2 (s)e2A(s) ds . f (t) ∼ N e−A(t) f0 + a(s)θ(s)eA(s) ds , e−A(t) 0
0
(3.19)
0
3.4 Lognormal processes The immediately natural extension of normal processes are all those that can be mapped back to a normal process. For instance, if we were interested in modelling a quantity S that is subject to a meanreversion effect, but cannot become negative, we could set f := ln S and describe the behaviour of S indirectly by saying that f satisfies equation (3.17). If we are then still interested in the explicit form of the stochastic differential equation governing the dynamics of S, we can apply Itˆo’s lemma to the inverse transformation S = ef to derive dS. This is exactly how we obtain the following stochastic differential equation for a mean-reverting lognormal process: dS 1 2 = a(t) [θ(t) − ln S] + σ (t) dt + σ(t)dWt S 2 The distribution of ln S is obviously given by (3.19).
(3.20)
3.5 The Markovian Wiener process embedding dimension
3.5
27
The Markovian Wiener process embedding dimension
Most systems of stochastic differential equations that we use for modelling of financial relevant quantities are of the from df = b(t, f )dt + A(t, f ) · dWt
(3.21)
with f , b ∈ Rn , Wt ∈ Rm , A ∈ Rn×m , and m ≤ n. The function b(t, f ) is frequently called the drift of the stochastic process, and the matrix A(t, f ) may be called the driving matrix or, following the nomenclature of Karatzas and Shreve [KS91], the dispersion matrix. Clearly, when viewed as embedded in the full state space dimensionality n, stochastic systems of the kind (3.21) are by definition Markovian, since none of the terms depend on the past or the future. Still, the reader may have come across the phrase that certain models are not Markovian, even though they can be described by an equation like (3.21). This slightly misleading notion usually arises when it is intended to express the fact that it is not possible to construct a recombining tree for f in the Wiener process dimensionality m. Of course, for all processes of the form (3.21) it is theoretically possible to construct a recombining tree when the embedding dimensionality of the tree is chosen to be n. However, one cannot in general reduce the embedding dimension of the tree to m. Whenever it is possible, though, the full system (3.21) may be called Markovian Wiener process embedding dimension reducible, or simply reducible. A process of the type of equation (3.21) is reducible if there is a bijection f (t) ⇐⇒ Wt .
(3.22)
In other words, given knowledge of the exact state of the m-dimensional driving Wiener process at any one time, it must be possible to identify exactly the values of all of the state variables in the vector f . This means that the process for f must not depend on the path that W took to get to its state at time t. Only then can we assign unique transition densities to each branch of the m-dimensional recombining tree1 . The above considerations arise naturally in the context of Heath-Jarrow-Morton models [HJM92a] of the yield curve. It can be shown that HJM models for the instantaneous forward rates f (t, T ) at time t for time T are reducible only for a very specific choice of the instantaneous volatility function σ(t, T ), and the specific choice leads to what is sometimes referred to as the generalised Hull-White model. Otherwise, the Markovian embedding dimension of any HJM forward rate model is infinite, even though only a few driving Wiener processes may be involved! Another yield curve modelling family that is not reducible is given by the Brace-Gatarek-Musiela/Jamshidian market models which always require as many dimensions as there are market rates in the model due to the specific state dependence of the BGM/J drift terms. For those problems, Monte Carlo methods are the numerical technique of choice since the difficulties involved with high-dimensionality are negligible for Monte Carlo methods, whilst for trees and lattice methods only a few dimensions are realistically tractable. 1
The discussion can be equally adapted to cater for a PDE formulation of the solution technique. It is only for the sake of clarity and convenience that we restrict the explanation to trees.
28
Stochastic dynamics
3.6
Bessel processes
A Bessel process Rt of dimension ν is given by the L2 -norm of a ν-dimensional Wiener process and probably owes its name to the sperical symmetry imposed by the L2 -norm: v u ν uX Rt := kWt k = t Wi2 t
(3.23)
i=1
It can be shown2 that a Bessel process of dimension ν satisfies the stochastic differential equation dRt =
ν − 1 dt + dWt 2 Rt
(3.24)
generated by a one-dimensional standard Wiener process dWt .
3.7 Constant elasticity of variance processes The constant elasticity of variance process [Bec80, Sch89] is given by df = σf γ dWt
for
γ≥0,
(3.25)
whereby I have omitted any drift terms. It has been used extensively throughout all areas of financial modelling, including equity, interest rates, commodities, and many other financial values. The reason that the CEV process is very popular is that it can reproduce a slightly curved implied volatility skew for plain vanilla options priced on the basis of it. For γ < 1, the skew is negative with options at low strikes having a higher implied volatility than those at high strikes. For γ > 1, the skew resulting from the CEV process is positive and describes the option prices observed in some commodity markets reasonably well. The formal solutions for the SDE (3.25) can be found in [BS96]. Alternatively, in the context of the BGM modelling framework for interest rates, a good summary of the properties of the CEV process and simple Call and Put option pricing formulas resulting from it can also be found in [AA00]. As for Monte Carlo simulations of the process (3.25), we will come back to the CEV process in section 4.3. The CEV process has a direct relation to the Bessel process. The transformation R :=
1 f 1−γ 1−γ
(3.26)
turns the SDE of the CEV process into that of a Bessel process of dimension ν=
1 − 2β . 1−β
(3.27)
This means, that for 0 ≤ β ≤ 1/2, the CEV process is equivalent to a Bessel process of a fractional dimension between 0 and 1, and can thus be absorbed at zero. For β > 1/2, however, the CEV process corresponds to a Bessel process of negative dimension and cannot attain zero. 2
See [KS91] page 159, equation (3.16).
29
3.8 Displaced diffusion
An interesting question in the context of section 3.5 is if a system of n equations of type (3.25) is reducible to the dimensionality m of the driving Wiener process Wt ∈ Rm . Clearly for γ = 0 or γ = 1, the process is reducible since we then have the standard normal or lognormal process, respectively. In general, however, as we discussed in section 3.5, for this to be the case, we must be able to write f (t) = F (Wt , t) for
f ∈ Rn
and W ∈ Rm
with
m≤n
for some function F (Wt , t). The application of Itˆo’s lemma to F then yields 1 2 dF = ∂W F dW + ∂t F + ∂W W F dt = df = f γ dW = F γ dW . 2
(3.28)
(3.29)
This means ∂W F = F γ
(3.30)
1 (1−γ) 1−γ + g(t) F = (1 − γ)W + F0
(3.31)
whose solution is
for some function g(t) with γ 6= 1 which we had already excluded anyway. The second condition to be satisfied by F that resulted from (3.28) is 1 ∂t F + ∂W W F = 0 2
(3.32)
1 g(t) ˙ = − γF 2γ−1 . 2
(3.33)
which, together with (3.31) implies
This last equation can only hold if g is a constant and γ = 0 because otherwise we would have a pure function of t on the left hand side, and a function of t and W on the right hand side which clearly is a contradiction. In other words, the CEV process is path-dependent, and a multi-dimensional CEV process driven by fewer Wiener processes than there are CEV state variables cannot be reduced.
3.8 Displaced diffusion Another process formulation that also gives rise to an implied volatility skew, in fact very similar to that of the CEV process, is the displaced diffusion process [Rub83] dS = µS dt + σdisplaced diffusion (S + θ) dW .
(3.34)
θ = − log2 (γ) · S0 ,
(3.35)
If we use the map
30
Stochastic dynamics
the displaced diffusion process can be used as a very good numerical proxy for the skew one would obtain from the CEV process, only with a somewhat more pronounced curvature (which is probably desirable). Unlike the CEV process, though, the displaced diffusion process is extremely easy to solve and numerically fast. Clearly we need to rescale the volatility similarly to how we translate from relative to absolute volatilities: S0 · σlognormal = (S0 + θ) · σdisplaced diffusion
(3.36)
As can be seen from (3.35), for γ = 1, the displaced diffusion process becomes ordinary geometric Brownian motion, i.e. a lognormal process. For γ =
1 2
the skew of the displaced diffusion process is
approximately that of the square root process, and in the limit γ → 0, we arrive at the normal evolution, similar to the CEV equation (3.25). The evolution of S subject to (3.34) out to a time horizon T is given by 1 2 ST = (S0 + θ) · e[(µ− 2 σdd )T +σdd
√
T z]
−θ
with
z ∼ N (0, 1) .
(3.37)
In the context of statistics, the distribution of ST is also known as a Johnson I distribution [JKB94, Joh49, Whe80].
Chapter 4 Process driven sampling In financial engineering and option pricing, we frequently wish to calculate the expectation of functionals of the time evolution of one or more underlying assets. In order to do this, one typically has to simulate a stochastic process and evaluate a pricing functional for each simulated path. In this chapter, some of the key concepts involved in this task are explained. Also, I will demonstrate how different methods work and recommend which technique is best for which class of problem.
4.1
Strong versus weak convergence
The numerical approximation of the solution of a stochastic differential equation amounts to a computing scheme that creates a sequence of numbers representing the time-discretisation of a specific sample path of a mathematically defined stochastic process. For ordinary differential equations, there is a natural meaning to the concept of convergence of a numerical method to the solution. For SDEs, the situation is different. The concept of strong convergence can be seen as the equivalent to convergence for ODEs since it requires that the numerical approximation to a given path of the stochastic process matches the truly exact solution at any point along the path. However, the solution to a stochastic differential equation at any time horizon, unlike ODEs, is not a single number but a (possibly multidimensional) distribution. And as for this distribution, we are often not interested in obtaining its exact functional shape but rather in the expectation of some function (typically a payoff profile) over that distribution. A time-discretised approximation Yδ of steps not larger than δ is said to be of general strong convergence order γ if for any time horizon T the approximation Yδ (T ) is guaranteed to converge to the exact solution X(T ) in expectation by absolute difference like E X(T ) − Yδ (T ) ≤ c δ γ
(4.1)
for all δ that are smaller than some positive δ0 and some constant c. In contrast to the strong convergence criterion by absolute difference, weak convergence of order β only requires that the numerically calculated expectation of any function g(·) which is 2(β + 1)31
32
Process driven sampling
times continuously differentiable and of polynomial growth converges to the exact equivalent. In other 2(β+1)
words, for any g ∈ CP
, E[g (X(T ))] − E[g (Yδ (T ))] ≤ c δ β
(4.2)
must hold for all δ that are smaller than some positive δ0 and some constant c. A consequence of the above definitions of strong and weak convergence is that many schemes of a given strong convergence order are of a higher weak convergence order. An example of this is the standard Euler scheme which, whilst being an order 1/2 strong scheme, is also an order 1 weak scheme, given suitable smoothness and growth conditions on the drift and diffusion coefficients. It is therefore not surprising that in general any method that is of higher convergence order for the strong criterion, also turns out to be of superior performance for the purpose of option pricing. Another point of interest is that, for many schemes, the convergence order, weak or strong, which they are guaranteed to have for any general SDE, may actually be exceeded for specific stochastic differential equations. In particular for the kind of SDEs that we encounter in finance, this is often the case.
4.2
Numerical solutions
Whilst there are a great many different kinds of numerical schemes for the solution of stochastic differential equations, and the best reference is almost certainly the book by Kloeden and Platen [KP99], we now briefly introduce and discuss the Euler and the Milstein scheme. The starting point is in both cases the following general form of a stochastic differential equation dX = a dt + b dW .
(4.3)
Note that both a and b can be functions of the process variable X and time. In the multi-dimensional case of m state variables Xi driven by d independent Wiener processes, we have dXi = ai (t, X) dt +
d X
bij (t, X) dWj .
(4.4)
j=1
4.2.1
The Euler scheme
Denote the numerical approximation to the solution of (4.3) for a scheme over equal steps of size ∆t at time n·∆t as Y (tn ). The Euler scheme is then given by Y (tn+1 ) = Y (tn ) + a(tn , Y (tn )) ∆t + b(tn , Y (tn )) ∆W .
(4.5)
The Euler scheme is of strong convergence order 1/2 which means we can always fall back to this workhorse of a numerical procedure to test any other method.
33
4.2 Numerical solutions
In financial applications, we are often interested to represent the evolution of an asset by a stochastic process. The most common assumption is probably that of geometric Brownian motion: dS = (r − d)dt + σdW S
.
In this case, the Euler scheme becomes n o S(tn+1 ) = S(tn ) · 1 + (r − d)∆t + σ∆W n o √ = S(tn ) · 1 + σz ∆t + (r − d)∆t
(4.6)
(4.7) with
z ∼ N (0, 1) .
In the multidimensional case, the Euler scheme is Yi (tn+1 ) = Yi (tn ) + ai (tn , Y (tn )) ∆t +
d X
bij (tn , Y (tn )) ∆Wj .
(4.8)
j=1
4.2.2
The Milstein scheme
The Milstein scheme involves the addition of the next order terms of the Itˆo-Taylor expansion of equation (4.3). This gives Y (tn+1 ) = Y (tn ) + a(tn , Y (tn )) ∆t + b(tn , Y (tn )) ∆W + 12 bb0 ∆W 2 − ∆t .
(4.9)
with b0 =
∂b(t, X) . ∂X
(4.10)
For the case of geometric Brownian, this results in S(tn+1 ) = S(tn ) · 1 + (r − d − 12 σ 2 )∆t + σ∆W + 12 σ 2 (∆W )2 n √ o = S(tn ) · 1 + σz ∆t + r − d + 12 σ 2 z 2 − 1 ∆t
(4.11)
Although the Milstein scheme is definitely manageable in the one-dimensional case, its general multidimensional extension is not as straightforward as one may expect. It requieres not only the drawing of standard normal variates for the simulation of the standard Wiener process increments ∆W for each dimension, but additional ones to account for the Itˆo integrals involving the mixing terms Pd j=1 bij (t, X) dWj .
4.2.3
Transformations
Let us have a closer look at the difference the Milstein scheme made in comparison to the Euler scheme for geometric Brownian motion. The additional terms of the Milstein scheme amount to adding 1 2 σ 2
∆W 2 − ∆t
34
Process driven sampling
to the terms in the braces on the right hand side of (4.7). If we compare the two schemes for geometric Brownian motion with the exact analytical solution 1
S(tn+1 ) = S(tn ) · e(r−d− 2 σ
2 )∆t+σ∆W
,
(4.12)
we find that the Milstein scheme essentially just added all the terms that make the scheme exact up to order O (∆t) in expectation and variance. We could have achieved almost the same effect by simply transforming the original equation for geometric Brownian motion (4.6) to logarithmic coordinates by application of Itˆo’s lemma to obtain: d ln S = (r − d − 12 σ 2 )dt + σdW .
(4.13)
This similarity is not coincidental. In fact, in particular for the equations we deal with in finance, it is almost always preferable to transform the original SDE to a more amenable form and then simply use the original Euler scheme. Let us now demonstrate the similarity between the Milstein scheme and a suitably chosen transformation using the example of a mean-reverting square-root process √ dv = a(θ − v)dt + λ v dW .
(4.14)
This is essentially the Cox-Ingersoll-Ross interest rate model [CIR85] if we interpret v as the short rate. This kind of process is also popular to describe the behaviour of stochastic volatility, see, e.g. [Hes93]. The Milstein discretisation scheme of (4.14) is ∆v = where we have substituted
√
√ √ a(θ − v) − 14 λ2 ∆t + λ v ∆t z + 14 λ2 ∆t z 2
(4.15)
∆t z with z ∼ N (0, 1) for the Wiener process increment ∆W . Now, let
us consider a generic function u of v, i.e. u = u(v). Itˆo’s lemma gives us the stochastic differential equation for u: 2 ∂u ∂u √ 1∂ u 2 du = a(θ − v) + 2 2 λ v dt + λ v dW ∂v ∂v ∂v
(4.16)
We should now make a fortuitous choice of u(v) so that the new equation is better behaved. This is done by ensuring that the term in front of the driving process becomes independent on the state variable. In other words, we choose √ −1 ∂u ∝ v ∂v
or specifically u =
√
v.
(4.17)
This gives us du =
1 a(θ − u2 ) − 14 λ2 dt + 12 λdW 2u
(4.18)
and we finally have the following Euler scheme for the transformed SDE: ∆u =
√ 1 a(θ − u2 ) − 14 λ2 ∆t + 12 λ ∆t z 2u
(4.19)
35
4.2 Numerical solutions
In order to compare the convergence order in v using the Euler scheme in the transformed equation and the Milstein scheme in the original equation, we now calculate (u + ∆uEuler )2 − u2 − ∆vMilstein
(v(u + ∆uEuler ) − v(u)) − ∆vMilstein =
(4.20)
= 2u∆uEuler + ∆u2 − ∆vMilstein √ 3 = 0 + O( ∆t ) . The above relationship between a transformation to an SDE whose stochastic term no longer contains the state variable and the Milstein scheme holds in general. Given the general SDE dv = a(t, v)dt + b(t, v)dW ,
(4.21)
the Milstein scheme is √ a − 12 b ∂v b ∆t + b ∆t z + 12 b ∂v b ∆t z 2
∆v =
(4.22)
with z ∼ N (0, 1). For a general transformation u = F (v) to a new state variable u and the inverse transformation v = G(u), the SDE (4.21) becomes du = F 0 a + 12 F 00 b2 dt + F 0 b dW
(4.23)
by virtue of Itˆo’s lemma. Given the choice F 0 (v) =
1 b(t, v)
which leads to
F 00 (v) = −
1 ∂v b , b2
(4.24)
the Euler scheme in the transformed variable reads ∆u =
a b
√ − 12 ∂v b ∆t + ∆t z .
(4.25)
In order to compare, we calculate the Taylor expansion of the inverse transformation v(u + ∆u) = v(u) + G0 (u)∆u + 12 G00 (u)∆u2 + O(∆u3 ) .
(4.26)
Since dv 1 = du du dv
,
i.e.
G0 (u) =
1 F 0 (G(u))
= b(t, G(u))
(4.27)
and thus G00 (u) = G0 ∂v b = b ∂v b ,
(4.28)
we have v(u + ∆u) − v(u) =
√ 3 a − 12 b ∂v b ∆t + b ∆t z + 12 b ∂v b ∆t z 2 + O(∆t 2 ) .
(4.29)
As we can see, the Euler scheme in the transformed equation leads to a procedure that is equal in convergence up to order O(∆t) (inclusive) to the Milstein scheme in the original variable. An additional benefit of transformations that remove all dependence of the multiplicative factor in front of the
36
Process driven sampling
driving Wiener process increments on the state variables is that the resulting equations become readily amenable to a particularly simple predictor-corrector scheme. Since this predictor corrector scheme is of weak convergence order 1 for the transformed variable, we typically obtain an integration scheme of weak convergence order higher than 1 for the original state variable without the need for a complicated integration algorithm that requires the draw of more normal variates than we have driving Wiener processes over any one time step.
4.2.4
Predictor-Corrector
One way to describe the idea behind predictor-corrector methods for stochastic differential equations is as follows. Given a time-discretised approximation to a general SDE such as equation (4.4), we know that taking an Euler step as in (4.8) ignores the fact that the coefficients ai and bij actually change along the path over the time step ∆t. Now, if that is so, wouldn’t it be better to use an approximate average value for those coefficients along the path over the finite time step ∆t? Since the values of the coefficients depend on the state variables themselves, which we don’t know, we need to approximate those first. The simplest predictor-corrector scheme, which, incidentally is of weak convergence order 1, is thus as follows: First take an Euler step as in equation (4.8) to arrive at the predictor Y¯i (tn+1 ) = Yi (tn ) + ai (tn , Y (tn )) ∆t +
d X
bij (tn , Y (tn )) ∆Wj .
(4.30)
j=1
Next, select two weighting coefficients α and η in the interval [0, 1], usually near 1/2, and calculate the corrector Yi (tn+1 ) = Yi (tn ) + α a ¯i (tn+1 , Y¯ (tn+1 ); η) + (1 − α)¯ ai (tn , Y (tn ); η) ∆t m X √ + η bij (tn+1 , Y¯ (tn+1 )) + (1 − η)bij (tn , Y (tn )) ∆t zj
(4.31)
j=1
with a ¯i (t, Y ; η) := ai (t, Y ) − η
m X d X
bkj (t, Y )∂Yk bij (t, Y ) .
(4.32)
j=1 k=1
Clearly, this scheme is very easy to implement, in particular for the special case that the coefficients bij don’t depend on the state variables.
4.3 Spurious paths Anyone who implements the straightforward Euler scheme (4.7) for geometric Brownian motion will notice a strange thing to occur every now and then: Some paths cross zero! Clearly, geometric Brownian motion should never even reach the point zero, let alone cross it. The reason why this happens is simple. The scheme (4.7) is only an approximation and only guaranteed to converge to the
37
4.4 Strong convergence for Euler and Milstein
mathematically consistent description of the geometric Wiener process in the limit of ever smaller time steps ∆t. In any simulation with a finite step size, it is only a matter of time until you draw a normal variate z that satisfies z x0 ∧ y < y 0 or x < x0 ∧ y > y 0 . Kendall’s tau is defined as the difference between the probabilities of two such pairs being concordant and discordant, i.e. τK := P [(x − x0 )(y − y 0 ) > 0] − P [(x − x0 )(y − y 0 ) < 0] . 2
provided they are not pathologically malicious
(5.11)
45
5.2 Copulæ
Naturally, this means that τK ∈ [−1, 1]. If we define the distributional densities and cumulative probabilities of x and y as in the previous section, and in addition Z Z x,y Ψ(x, y) := ψ(x0 , y 0 ) dx0 dy 0 ,
(5.12)
x0 ,y 0 =−∞
it can be shown that an alternative formula for Kendall’s tau for continuous distribution densities is ZZ Ψ(x, y)ψ(x, y) dxdy − 1 . (5.13) τK = 4 Since Kendall’s tau is defined on the joint cumulative probability, it is also invariant with respect to transformations3 . Kendall’s tau and Spearman’s rho belong to the category of rank correlations. Rank correlations have the nice property that for any two marginal distribution densities ψx (x) and ψy (y), there always exists a joint distribution density ψ(x, y) for every possible value in [−1, 1] of the rank correlation.
5.2
Copulæ A copula of two variables x and y is a cumulative probability function defined directly as a function of the marginal cumulative probabil-
ities of x and y. A copula is thus a way to specify the co-dependence between two variates entirely independently on their individual marginal distribution. By this definition, a copula of n variables is a function C : [0, 1]n → [0, 1]. Reusing the definitions (5.5), (5.6), and (5.12) we can thus identify Ψ(x, y) = C (Ψx (x), Ψy (y)) .
(5.14)
For strictly increasing cumulative marginals Ψx (x) and Ψy (y), we can also write −1 C(u, v) = Ψ Ψ−1 x (u), Ψy (v) .
(5.15)
The copula of independent variables, not surprisingly, is given by Cindependent (u, v) = u · v.
(5.16)
By virtue of the definition on the cumulative marginal distribution functions, the copula of a set of variables (x, y) is invariant with respect to a set of strictly increasing transformations (f (x), g(y)). The differential of a copula is sometimes written by the notation dC(u, v) which is to mean ∂(u, v) −1 −1 dudv dC(u, v) = ψ Ψx (u), Ψy (v) ∂(x, y) 3
(5.17)
again: as long as they are reasonably benign
Here is a puzzle for you: What does the following piece of code do and how is it connected to the excerpt from a dictionary at the beginning of section 5.2 ? /d { rand 2147483647. div } def /r { d d d add add } def /normal { r r r r add add add 6 sub } def /T 415 def /dt 1 def /sigma .5 def /a -.25 def 0 0 dt T { pop dup a mul dt mul sigma dt sqrt mul normal mul add dup dt exch rlineto add } for
46
Correlation and co-movement
with u = Ψx (x) and v = Ψy (y). In this notation, Kendall’s tau appears as ZZ τK = 4 C(u, v) dC(u, v) − 1 .
(5.18)
[0,1]2
Naturally, all of the above definitions extend to more than just two variables. Note that a sensible copula must be a function of all the cumulative marginal probabilities, and some kind of control parameters that determine the strength of co-dependence of the individual variables.
5.2.1
The Gaussian copula
The short explanation of the Gaussian copula mechanism is: Generate a vector of correlated normal variates, transform them back to uniform variates by the aid of the cumulative normal probability function, and then map the variates into their individual marginal distributions using their respective inverse cumulative marginal probability functions. In other words, if we wish to construct the variate vector x subject to a vector of individual marginal distribution densities ψ x (x) = (ψx1 (x1 ), ψx2 (x2 ), ..., ψxn (xn ))> and coupled by the Gaussian copula controlled by the correlation matrix R, we need to proceed as follows: · Find a suitable pseudo-square root A of R such that R = A · A> . More on this on page 58 in chapter 6. · Draw a vector z ∈ Rn of uncorrelated standard normal variates. ˜ := A · z. · Compute z ˜ back to a vector of uniform variates v ∈ [0, 1]n by setting vi = N(˜ · Map z zi ). · Construct the variate vector x by use of the inverse cumulative probability functions xi := Ψ−1 xi (vi ). It is important to remember that the correlation coefficient governing the Gaussian copula can be very different from the linear correlation of the elements of the variate vector x. I give below two specific examples for this effect. First, however, there is one further point to mention. It can be shown [LMS01, Kau01] that Kendall’s tau of two variables connected by a Gaussian copula with correlation coefficient ρ is given by τK =
2 arcsin ρ . π
(5.19)
Examples for the density of the Gaussian copula as defined in (5.17) are shown 4 in figure 5.2. 4
The colour code in this and the subsequent density diagrams, although only visible in the electronic version of this book on the accompanying CD, is as follows: red signifies the lowest density, i.e. 0, and purple the highest density which is usually only a tiny little speck in the diagram. In all figures, the colours are nonlinearly scaled with the density such that centre point at (0.5, 0.5) always has the same turquois colour throughout all of the density figures.
47
5.2 Copulæ Density of Gaussian copula for ρ= 0.2
Density of Gaussian copula for ρ= 0.7
Density of Gaussian copula for ρ= 0.99
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
0 0
0.2
0.4
0.6
0.8
1
0 0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
Figure 5.2: The Gaussian copula density as given by equation (5.17). For ρ = 0, the density is identically equal to 1 on the unit square. The density for negative values of ρ corresponds to the density for the same absolute value by a rotation of 90◦ . Two uniform variates under the Gaussian copula Let us assume that for some purpose we would like to have two uniform variates that are correlated. Our first thought of constructing correlated uniforms is to transform two uncorrelated uniform variates u1 and u2 to the standard normal distribution by setting zi = N−1 (ui ), apply the conventional correlation transformation z˜1 = z1
(5.20)
p z˜2 = ρz1 + 1 − ρ2 z2
(5.21)
for some correlation coefficient ρ ∈ [−1, 1] and transform back to the unit square by the aid of u˜i = N(˜ zi ). Na¨ıvely, we now expect the co-dependent uniform variates to have a linear correlation coefficient in the range [−1, 1]. But little do we know... The linear correlation η between the two dependent uniform variates can be calculated: ZZ p η(ρ) = 12 N(z1 )N(ρz1 + 1 − ρ2 z2 )ϕ(z1 )ϕ(z2 ) dz1 dz2 − 3 .
(5.22)
Straightforward calculus shows that η ∈ [−1, 1] for
ρ ∈ [−1, 1] .
Near the origin, we have η(ρ) ≈
3 ρ π
for
|ρ| 1 .
(5.23)
A diagram of η(ρ) is shown in figure 5.3. Don’t be misled by the apparently straight line: there is a little bit of curvature in there, although a straight line would certainly be a good approximation for it.
1
48
Correlation and co-movement 1
η(ρ)
0.5
0
-0.5
-1 -1
-0.5
0 ρ
0.5
1
Figure 5.3: The linear correlation η(ρ) of two uniform variates connected by a Gaussian copula of correlation coefficient ρ. Two exponential variates under the Gaussian copula In certain areas in finance, we are interested in the modelling of the time elapsed until a certain event happens for the first time. When we model a given event A using a Poisson process of intensity λA as explained in section 2.3 we have the following distribution density for the time TA until arrival: TA ∼ λA e−λA TA .
(5.24)
Random draws for TA can be generated from a uniform variate uA by setting TA = −
ln(1 − uA ) . λA
(5.25)
When we model the co-dependent arrival of two events A and B using a Gaussian copula with correlation coefficient ρ, we can calculate the linear correlation ζ(ρ) for TA and TB as ZZ p ζ(ρ) = ln(1 − N(z1 )) ln(1 − N(ρz1 + 1 − ρ2 z2 ))ϕ(z1 )ϕ(z2 ) dz1 dz2 − 1 .
(5.26)
Again, we can evaluate analytically what interval ζ(ρ) is confined to: ζ ∈ [1 − where 1 −
π2 6
π2 , 1] for 6
ρ ∈ [−1, 1] ,
≈ −0.6449341. However, as we can see in figure 5.4, the correlation transformation is
not quite as nearly linear as it was for two uniform variates.
49
5.2 Copulæ 1 0.8 0.6
ζ(ρ)
0.4 0.2 0 -0.2 -0.4 -0.6 -1
-0.5
0 ρ
0.5
1
Figure 5.4: The linear correlation ζ(ρ) of two exponential variates connected by a Gaussian copula of correlation coefficient ρ.
5.2.2 The t-copula The t-copula is conceptually very similar to the Gaussian copula. It is given by the cumulative distribution function of the marginals of correlated t-variates. The simplest way to explain the t-copula is probably by the aid of an algorithm that would create uniform variates under a t-copula [ELM01]. · Select a standard correlation matrix R that is to govern the co-dependence of the copula. · Find a suitable pseudo-square root A of R such that R = A · A> . · Draw a vector z ∈ Rn of uncorrelated standard normal variates. ˜ := A · z. · Compute z · Draw an independent χ2ν -variate s. For ν being an integer, this can be done by drawing ν independent Gaussians, and summing their squares. · Set x :=
pν s
˜. z
· Map x back to a vector of uniform variates v ∈ [0, 1]n using the cumulative probability function of Student’s t distribution given in equation (2.52).
50
Correlation and co-movement
The so generated uniform variates can then be transformed to any set of marginal distributions by use of their respective inverse cumulative probability functions, just as for the Gaussian copula. The t-copula shares with the Gaussian copula the feature that two variates, which are connected using either of the two copulæ with a given correlation coefficient ρ, have a Kendall’s tau coefficient given by equation 4
4
3
3
2
2
1
1
0
0
y
y
(5.19). An example for the difference from both copulæ is shown in figure 5.5 Note how the t-copula
-1
-1
-2
-2
-3
-3
-4
-4 -4
-3
-2
-1
0 x
1
2
3
4
-4
-3
-2
-1
0 x
1
2
3
4
Figure 5.5: Both figures show data that are marginally normally distributed. The figure on the left depicts ordinary correlated normal variates with ρ = 0.9, and the figure on the right was created from data under a t2 -copula, also with ρ = 0.9. generates apparently higher correlation for large co-movements of equal sign, at the expense of giving rise to a higher density near the origin, and a noticeable set of what looks like significant outliers.
5.2.3
Archimedean copulæ
All members of this class of copulæ have in common that they are generated by a strictly decreasing convex function φ(u) which maps the interval (0, 1] onto [0, ∞) such that lim→0 φ() = ∞ and φ(1) = 0. An Archimedean copula is generated from a given function φ(u) by C(u, v) = φ−1 (φ(u) + φ(v)) .
(5.27)
Two uniform variates u and v under any Archimedean copula can be produced by the following algorithm: · Draw two independent uniform variates s and q. · Solve q =t−
φ(t) φ0 (t)
(5.28)
51
5.2 Copulæ for t. · Set u := φ−1 (sφ(t))
(5.29)
v := φ−1 ((1 − s)φ(t)) .
(5.30)
For further details and a proof see [ELM01]. An example of a copula generator for which the above algorithm can be applied directly is θ φ(u) = u−1 − 1
(5.31)
for θ ≥ 1 since then the solution to equation (5.28) is immediately given by s 2 1+θ 1+θ t= − − θq . 2 2
(5.32)
It is possible to extend Archimedean copulæ to higher dimensions. However, these extensions are rather restrictive with respect to the co-dependence structure since they do not allow for one parameter per pair of variates, unlike the Gaussian and the t-copula. This is probably the major disadvantage of Archimedean copulæ. For algorithms on the generation of variates under a higher dimensional Archimedean copula see [ELM01, FV97]. The Gumbel copula The Gumbel copula (sometimes also referred to as Gumbel-Hougaard copula) is controlled by a single parameter θ ∈ [1, ∞). It is generated by φGumbel (u) = (− ln u)θ
(5.33)
and thus defined as CGumbel (u, v) = e−[(− ln u)
1 θ +(− ln v)θ θ
] .
(5.34)
The Gumbel copula gives rise to a stronger dependence in the upper tail of the joint distribution density as can be seen in figure 5.6. Kendall’s tau of the Gumbel copula can be shown [GR00] to be τGumbel = 1 −
1 . θ
(5.35)
The Clayton copula This copula is also known under the names Pareto, Cook-Johnson, or Oakes copula, and is generated by the definition (u−θ − 1) φClayton (u) = θ
(5.36)
52
Correlation and co-movement Density of Gumbel copula for θ = 1.1
Density of Gumbel copula for θ = 1.5
Density of Gumbel copula for θ = 2
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
0 0
0.2 0.4 0.6 0.8 Density of Gumbel copula for θ = 3
1
0 0
0.2 0.4 0.6 0.8 Density of Gumbel copula for θ = 5
1
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
0 0
0.2
0.4
0.6
0.8
1
0
0.2 0.4 0.6 0.8 Density of Gumbel copula for θ = 10
1
0
0.2
1
0 0
0.2
0.4
0.6
0.8
1
0.4
0.6
0.8
Figure 5.6: The Gumbel copula density as given by equation (5.17). For θ = 1, the density is identically equal to 1 on the unit square. fior θ ∈ [−1, ∞) \ {0} and reads − θ1 −θ −θ CClayton (u, v) = max u + v − 1 ,0 .
(5.37)
The Clayton copula gives rise to a stronger dependence in the lower tail of the joint distribution density as can be seen in figure 5.7. Kendall’s tau of the Clayton copula is τClayton =
θ . θ+2
(5.38)
The Frank copula The Frank copula is given by (e−θu − 1) φFrank (u) = − ln (e−θ − 1)
(5.39)
and ! e−θu − 1 e−θv − 1 1 CFrank (u, v) = − ln 1 + θ (e−θ − 1)
(5.40)
53
5.2 Copulæ Density of Clayton copula for θ = 0.1
Density of Clayton copula for θ = 1
Density of Clayton copula for θ = 2
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
0 0
0.2 0.4 0.6 0.8 Density of Clayton copula for θ = 3
1
0 0
0.2 0.4 0.6 0.8 Density of Clayton copula for θ = 5
1
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
0 0
0.2
0.4
0.6
0.8
1
0
0.2 0.4 0.6 0.8 Density of Clayton copula for θ = 10
1
0
0.2
1
0 0
0.2
0.4
0.6
0.8
1
0.4
0.6
0.8
Figure 5.7: The Clayton copula density as given by equation (5.17). For θ = 0, the density is identically equal to 1 on the unit square. for θ ∈ R \ {0}. The Frank copula enhances the upper and lower tail dependence equally as can be seen in figure 5.8. For negative values of θ, this copula is able to produce negative codependence similar to the Gaussian copula for negative ρ. Not only that, but it also displays the invariances with respect to (u, v) → (1 − u, 1 − v) and (u, v; θ) → (1 − u, v; −θ), as does the Gaussian copula, which is demonstrated in figure 5.8. For the Frank copula, Kendall’s tau is (1 − D1 (θ)) θ
τFrank = 1 − 4
(5.41)
with Dk (x) for some positive integer k being the Debye function [AS84, Mac96] defined as k Dk (x) = k x
Z 0
x
sk ds . es − 1
(5.42)
The Ali-Mikhail-Haq copula This copula is generated by [AMH78] φAli-Mikhail-Haq (u) = ln
1 − θ(1 − u) u
(5.43)
54
Correlation and co-movement Density of Frank copula for θ = − 1
Density of Frank copula for θ=−5
Density of Frank copula for θ=−10
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
0 0
0.2 0.4 0.6 0.8 Density of Frank copula for θ = 1
1
0 0
0.2 0.4 0.6 0.8 Density of Frank copula for θ = 5
1
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
0 0
0.2
0.4
0.6
0.8
1
0
0.2 0.4 0.6 0.8 Density of Frank copula for θ = 10
1
0
0.2
1
0 0
0.2
0.4
0.6
0.8
1
0.4
0.6
0.8
Figure 5.8: The Frank copula density as given by equation (5.17). For θ → ±0, the density approaches the uniform density on the unit square. and has the form CAli-Mikhail-Haq (u, v) =
uv 1 − θ(1 − u)(1 − v)
(5.44)
for θ ∈ [−1, 1]. The Ali-Mikhail-Haq copula enhances lower tail dependence for positive θ, and displays some strong negative codependence for θ < 0 as is shown in figure 5.9. Kendall’s tau is [FV97] τAli-Mikhail-Haq =
3θ − 2 θ
2 − 3
1 1− θ
2 · ln(1 − θ) .
(5.45)
A generalisation of the Ali-Mikhail-Haq copula to two parameters known as the Fang-Fang-Rosen copula [GR00, FFvR00] is given by 1
φFang-Fang-Rosen (u) = ln
1 − θ(1 − u κ ) 1
uκ
! .
(5.46)
55
5.2 Copulæ
Density of Ali− Mikhail− Haq copula for θ = − 1
Density of Ali− Mikhail− Haq copula for θ = − 0.3
Density of Ali− Mikhail− Haq copula for θ = 0.001
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
0 0
0.2 0.4 0.6 0.8 Density of Ali− Mikhail− Haq copula for θ = 0.3
1
0 0
0.2 0.4 0.6 0.8 Density of Ali− Mikhail− Haq copula for θ = 0.6
1
0
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
0 0
0.2
0.4
0.6
0.8
1
0.2 0.4 0.6 0.8 Density of Ali− Mikhail− Haq copula for θ = 0.99
1
0 0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
Figure 5.9: The Ali-Mikhail-Haq copula density as given by equation (5.17). For θ = 0, the density is uniformly 1 on the unit square.
1
56
Correlation and co-movement
Chapter 6 Salvaging a linear correlation matrix The problem of how to specify a correlation matrix occurs in several important areas of finance. A few of the important applications are the specification of a (possibly time-dependent) instantaneous correlation matrix in the context of the BGM interest-rate option models, stress-testing and scenario analysis for market risk management purposes, or the specification of a correlation matrix amongst a large number of obligors for credit-derivative pricing or credit risk management. Ad-hoc correlation matrices, those calculated from incomplete data, and those taken from news services sometimes don’t comply with the requirement of symmetry and positive semi-definiteness. Whilst it is easy to amend the symmetry requirement by manual intervention, it is not always straightforward to see how to adjust the given correlation matrix to become usable for factor analysis or simulation purposes. What’s more, there are many situations when it is desirable to carry out a calculation not only for a single input matrix, but for a whole set of modified versions of the original matrix. Examples for this include comparative pricing in order to ascertain the extent of correlation exposure for multi-asset derivatives, but also the assessment of portfolio risk. In many of these cases, we end up with a matrix that is no longer positive semi-definite, and often there is no clear way to remedy this. In practice, the problem of an invalid correlation matrix, i.e. one that has negative eigenvalues, can also very easily arise in the context of risk analysis for equity portfolios. This is because there are frequently asynchronous gaps in the historical stock exchange time series. The chance that slight inconsistencies in the data from which historical correlation coefficients are calculated can lead to negative eigenvalues grows rapidly as the size of the correlation matrix increases. This has recently been pointed out by Ju and Pearson [XP99]. Intuitively, it can be understood to be an effect of the characteristic polynomial that determines the eigenvalues becoming of higher order as the dimension of the correlation matrix grows, and thus displaying a stronger nonlinear response to slight changes in the polynomial’s coefficients. Since equity index or portfolio analysis typically involves many underlying assets, the risk of negative eigenvalues of the correlation matrix calculated from historical data is particularly large. In this chapter, I describe two methods based solely on mathematical grounds which can be used to best-match an invalid correlation matrix given the constraint of positive-semidefiniteness. Not only 57
58
Salvaging a linear correlation matrix
are these guaranteed to give a solution, but in addition we also have a measure to what extent we are matching the target matrix. Both methods are centered around the idea of decomposing a covariance matrix C into its pseudo-square root which is to mean any matrix A such that C = A · A> . For scalars, A would be an actual square root of C, but for matrices, the concept of a square root is ill defined whence I use the term pseudo-square root.
6.1
Hypersphere decomposition
The starting point are the well known results from linear algebra that every n × n matrix M given by M = WW>
(6.1)
for any W ∈ R n×n is positive-semidefinite and that, conversely, every positive-semidefinite matrix M ∈ R n×n can be decomposed as in equation (6.1). The hypersphere decomposition method for the construction of a valid correlation matrix ˆ = BB > C
(6.2)
that best-matches a given, not positive-semidefinite, target matrix C is to view the elements of the row vectors of matrix B in equation (6.2) as coordinates lying on a unit hypersphere [RJ00]. If we denote by bij the elements of the matrix B, the key is to obtain the n × n coordinates bij from n × (n − 1) angular coordinates θij according to bij
=
cos θij ·
j−1 Y
sin θik
for
j = 1 .. n − 1
k=1
and
(6.3) j−1
bij
=
Y
sin θik
for
j=n
.
k=1
ˆ formed from B as in equation (6.2) satisfies all the given For an arbitrary set of angles {θij }, a matrix C constraints required of a correlation matrix by construction. In particular, thanks to the trigonometric relationship (6.3) and to the fact that the radius of a unit hypersphere is always equal to one, the main diagonal elements are guaranteed to be unity which is shown in section 6.5. ˆ will bear no resemblance to the target matrix C. However, after using the In general, matrix C above transformation and after defining a suitable error measure ε in the resulting approximate correlˆ ation matrix C ˆ ε = kC − Ck
,
(6.4)
one can use an optimisation procedure over the angles θij to find the best possible fit given the chosen error measure. Sensible choices for the error measure are:-
59
6.2 Spectral decomposition
ˆ • the sum of squares of the elements of the difference matrix C − C , χ2Elements :=
X
(cij − cˆij )2
.
(6.5)
ij
ˆ have unit diagonal elements, this error norm is equal to twice the sum of Since both C and C squares of errors in the free correlation coefficients. ˆ • the elementwise sum of squares of errors in the sorted sets of eigenvalues of C and C, χ2Eigenvalues :=
X
ˆi λi − λ
2
.
(6.6)
i
Naturally, the above suggestions are only examples and various other choices are conceivable. If, in particular, a risk manager felt that certain portions of the target correlation matrix C should be recovered with particularly high accuracy, then correspondingly large weights could be assigned to the relative elements (cij − cˆij )2 . The fundamental benefits of this method are twofold: first, when the underlying space over which the optimisation is carried out is expressed in terms of angle vectors describing coordinates on a unit hypersphere, no constraints have to be satisfied. Secondly, the approach presented in the next section requires no iterations and provides a solution very similar to the one obtained using error metric (6.5). It can therefore be used to provide the starting point for the search procedure.
6.2 Spectral decomposition This method is based on the idea that the spectrum, i.e. the set of eigenvalues, is the most significant criterion to be preserved in a matrix when we amend it to meet certain constraints1 . It is a particularly useful approach for the given problem since the violated constraint itself is that an eigenvalue is negative. Given the right-hand-side eigensystem S of the real and symmetric matrix C and its associated set of eigenvalues2 {λi } such that C ·S =S·Λ
with
Λ = diag (λi )
define the non-zero elements of the diagonal matrix Λ0 as ( λ i : λi ≥ 0 Λ0 : λ0i = 0 : λi < 0 1
.
,
(6.7)
(6.8)
This method is also known as principal component analysis. The combination of Householder reduction to tridiagonal form and the QL algorithm with implicit shifts for tridiagonal matrices provides a very efficient way of computing the eigenvalues and eigenvectors of real symmetric matrices. The Numerical Recipes [PTVF92] routines tred2, tqli, and eigsrt can be used in that order to carry out the decomposition. 2
60
Salvaging a linear correlation matrix
If the target matrix C is not positive-semidefinite, it has at least one negative eigenvalue whence at least one of the λ0i will be zero. Also, define the non-zero elements of the diagonal scaling matrix T with respect to the eigensystem S by #−1
" T
:
ti =
X
s2im λ0m
.
(6.9)
m
Now, let3 √ B 0 := S Λ0
(6.10)
and √ B :=
T B0 =
√
√ T S Λ0
.
(6.11)
For normalised row vectors of S, the truncation of the negative eigenvalues results in row vectors of B 0 that are not of unit length. This is rectified in equation (6.11) by the aid of matrix T which contains the required normalisation factors. By construction, ˆ := BB > C
(6.12)
is now both positive-semidefinite and has unit diagonal elements since its elements are cˆij =
√ √ X √ √ ( T )ik · skl · ( Λ0 )lm · ( Λ0 )mn · spn · ( T )pj klmnp
=
X√
ti · sil · λ0l · sjl ·
p
tj
l
P
sil sjl λ0l
l
= qP P s2im λ0m · s2jk λ0k m
.
(6.13)
k
A procedural description of the above method may clarify what actually has to be done: · Calculate the eigenvalues λi and the right-hand-side eigenvectors si of C. · Set all negative λi to zero. · Multiply the column vectors si with the square roots of their associated corrected eigenvalues λ0i and arrange them as the columns of B 0 . · Finally, B results from B 0 by normalising the row vectors of B 0 to unit length. √ Please note that the notation D for a diagonal matrix D with non-negative elements is a symbolic description of the diagonal matrix whose non-zero elements are the positive roots of the diagonal elements of D. 3
61
6.3 Angular decomposition of lower triangular form
By following this procedure we obtain an acceptable correlation matrix which is intuitively similar to the target one (the more so, the fewer the eigenvalues which have to be set to zero). The crucial point, however, is not so much the plausibility of the metric but the fact that empirically I have always observed the results obtained using equations (6.7) to (6.12) to be very similar to those from the angular method discussed in section 6.1. How close the results are in practice is shown in section 6.4. This is significant because one can use the result of the method described here either as an accurate approximation to the best (in a χ2Elements sense) solution, or as the starting point for the optimisation discussed in section 6.1, thereby substantially reducing the computational burden of the hypersphere decomposition approach.
6.3
Angular decomposition of lower triangular form
ˆ The form given in equation (6.3) is the most general decomposition B of any valid correlation matrix C ˆ = BB > . However, any so derived matrix B can be transformed by a generic orthogonal such that C matrix O without any change of the effective correlation matrix given by the product BB > . The group of all possible orthogonal matrices O ∈ Rn×n represents all possible rotations and reflections. Ignoring the latter and just taking into account all possible rotation matrices, we end up with
n×(n−1) 2
degrees of freedom in the rotation matrices given by the same number of rotation angles. By virtue of these degrees of freedom, we can rotate every decomposition matrix B such that the transformed decomposition matrix B 0 := BO is of lower triangular form. We can thus, without loss of generality, ˆ formulate the following reduced form for the hypersphere decomposition B 0 of C: b011 b0ij
= =
1, j−1 Y
sin θik · cos θij
for
j = 1 .. (i − 1) ,
sin θik
for
j=i,
k=1
b0ij
=
j−1 Y
(6.14)
k=1
and b0ij
=
0
for j = (i + 1) .. n
The above reduced form is identical with (6.3) if we choose θij = 0 for all j ≥ i. In matrix form, the lower triangular decomposition (6.14) thus looks as follows: 1 0 0 0 cos θ21 sin θ21 0 0 cos θ 0 sin θ31 sin θ32 0 B = 31 sin θ31 cos θ32 cos θ41 sin θ41 cos θ42 sin θ41 sin θ42 cos θ43 sin θ41 sin θ42 sin θ43 .. .. .. .. . . . .
0 ···
0 ··· 0 ··· 0 ··· .. . . . .
(6.15)
62
Salvaging a linear correlation matrix
One of the advantages of the completely general decomposition (6.3) was that the result of the spectral decomposition and truncation (6.11) could directly be used in a bootstrapping procedure to calculate good initial values for the angles θij for the subsequent optimisation procedure. In order to start off the optimisation in the case of the reduced form, the result of the spectral decomposition and truncaˆ as given in tion can still be used as follows. First, we re-constitute the effective correlation matrix C equation (6.12) from the spectrally truncated approximation (6.11). Then, we calculate the Cholesky ˆ (which can be done very efficiently [PTVF92]). Since the Cholesky decomposdecomposition of C ition is of lower triangular form, it can then be used directly to back out good initial guesses for the reduced number of
n×(n−1) 2
angles that specify the matrix B 0 as given in equations (6.14) and (6.15),
respectively. The advantage of the reduction of the number of free parameters in the optimisation procedure by 50% is obvious: halving the dimension of the optimisation domain decreases the number of entries in any involved Jacobian or Hessian matrix by a factor of four which should give rise to a speed-up factor of the same order of magnitude.
6.4 Examples A risk manager retrieves from the middle office’s reporting system the following correlation matrix of three world equity indices:
0.9 0.7 ˜ = C 0.9 1 0.4 0.7 0.4 1 1
.
˜ are { 2.35364, 0.616017, 0.0303474 } and the correlation matrix can be The eigenvalues of C split up as in ˜ =B ˜B ˜> C with
0.08718 −0.13192
0.98742
˜ = B 0.88465
0.45536
0.77203 −0.63329
0.10021 0.05389
.
The risk manager is aware of the Value at Risk calculated under the assumption of this correlation between the three indices. In order to assess the change in Value at Risk resulting from a decrease in correlation between two of the three underlying variables, the risk manager wishes to adjust the matrix to
0.9 0.7 C = 0.9 1 0.3 0.7 0.3 1 1
.
63
6.5 Angular coordinates on a hypersphere of unit radius
Unfortunately, the eigenvalues of C 0 are now { 2.29673, 0.710625, −0.00735244 }, and despite its plausible appearance, matrix C 0 is no longer an acceptable correlation matrix. This highlights how a minor change can lead to the violation of the requirement of positive-semidefiniteness of a correlation matrix. The system will now fail when trying to construct a split-up matrix B for the purpose of Monte Carlo simulations4 in order to calculate the Value at Risk under the new assumptions. Using the method outlined in section 6.1 with the error measure chosen to be χ2Elements as given by equation (6.5), we can calculate
0.99804
0.06265 0
ˆ = B 0.50209 0 0.86482 0.74020 −0.67239 0 with
1
0.89458
ˆ =B ˆB ˆ> = C 0.89458 0.69662
0.69662
1
0.30254 1
0.30254
and a total error of χ2Elements = 0.946 · 10−4 . In comparison, the method outlined in section 6.2 above, yields 0.99805 0.06238 0 ˆ = B 0.50292 0 0.86434 0.73974 −0.67290 0 to give us
1
0.89402
ˆB ˆ> = ˆ =B C 0.89402 0.69632
1 0.30100
0.69632
0.30100 1
.
One can notice that not only the total error of χ2Elements = 1.0 · 10−4 but also the individual elements are remarkably close to the values obtained by optimisation. Despite the fact that there is in general no guarantee that the results of the two methods are as close together as in this example, I have always found very good agreement between the two approaches.
6.5 Angular coordinates on a hypersphere of unit radius The i-th row vector of B as specified by equation (6.3) is given by (bi1 , bi2 , . . . , bi n−1 , bin ) = (cos θi1 , sin θi1 cos θi2 , . . . , 4
Qn−2 k=1
sin θik cos θi n−1 ,
(6.16) Qn−2 k=1
sin θik sin θi n−1 )
.
Recall that the construction of correlated normal variates from a vector of uncorrelated normal variates z is done by the transformation x = B · z with C = BB > .
64
Salvaging a linear correlation matrix
The sum of the squares
j=n P
b2ij will definitely be unity if the recursive relation
j=1 k=n X
sin2 θij cos2 θij
(6.17)
sin2 θi1 sin2 θi1 2 = cos θ · = sin2 θi1 i1 cos2 θi1 cos2 θi1
(6.18)
b2ik = b2ij ·
k=j+1
holds since then k=n X
b2ik = b2i1 ·
k=2
and thus k=n X
b2ik
b2i1
=
+
k=n X
k=1
b2ik = cos2 θi1 + sin2 θi1 = 1 .
(6.19)
k=2
To start the induction, we see that equation (6.17) is satisfied for j = n − 1 since then we have n−2 Y
2
2
sin θik sin θi n−1 =
k=1
n−2 Y
sin2 θik cos2 θi n−1 ·
k=1
sin2 θi n−1 cos2 θi n−1
.
(6.20)
It remains to be shown that b2ij
k=n cos2 θij X 2 · b = sin2 θij k=j+1 ik
(6.21)
for j < n − 1. Using the recursion relation (6.17) itself, we obtain ! k=n X cos2 θij 2 2 2 bij = · bi j+1 + bik sin2 θij k=j+2 cos2 θij sin2 θi j+1 2 2 = · bi j+1 + ·b cos2 θi j+1 i j+1 sin2 θij cos2 θij = · b2i j+1 , 2 2 sin θij cos θi j+1 i.e. bi j+1 = bij ·
sin θij cos θi j+1 cos θij
(6.22)
which is identical to the construction description (6.3) for j < n − 1. Hence, all row vectors of B ˆ = BB > are the pairwise scalar products of the row vectors are of unit length. The elements of C ˆ satisfies the of B. Since the scalar product of two vectors of unit length is by definition ∈ [−1, 1], C requirements of unit diagonal elements, symmetry, and all elements being ∈ [−1, 1].
Chapter 7 Pseudo-random numbers For all Monte Carlo methods, we need an underlying number generator. This driving engine has to supply us with variate vectors which in the limit of infinitely many draws satisfy a given joint multivariate distribution density function. This is typically done by transformation of draws from the uniform distribution of equal probability for all numbers in the interval (0, 1). Note that, unlike most textbook definitions of numerically generated uniform numbers, both 0 and 1 are explicitly excluded since for most of the desired target distributions at least one of the two endpoints maps to either +∞ or −∞ which clearly poses a numerical problem. Traditionally, Monte Carlo techniques used to depend on a number generation method that mimics randomness as well as possible and a great deal of effort has gone into number theoretical research for this purpose. Generations of number theoreticians have focussed on ever more refined and intricate ways of constructing random numbers whilst others devised ever more sophisticated tests for randomness [Knu81]. The reason for all of this hard work is that a machine that is designed to follow instructions in a deterministic way such as a computer cannot produce anything that actually is random. This was beautifully expressed by John von Neumann in his statement which has become known as ‘the original sin of random number generation’: Anyone who considers arithmetical methods of producing random digits is, of course, in a state of sin.
John von Neumann, 1951 [vN51] A more mathematical way to express this fundamental failure of randomness of computer-generated digits is that due to the algebraic nature of their generation, there always exists a high-dimensional embedding space Rd such that vector draws v whose elements are sequential draws from a onedimensional number generation engine can appear as systematically aligned in a lower dimensional manifold. One example for such a high-dimensional embedding is clearly given by the periodicity of the number generator which is why modern methods pay great attention to this feature and achieve very long periodicities. However, this continuous chase for ever more random numbers is somehow 65
66
Pseudo-random numbers
doomed because for any new method that satisfies all known tests, one can always construe a new test criterion for randomness which will prove it to be non-random: Every random number generator will fail in at least one application.
Donald E. Knuth, 1969 It is for this reason that computer-generated random numbers are referred to as pseudo-random numbers [Sob94]. They simply cannot be random. Recently, there have been several attempts to overcome the very last bit of non-randomness left in modern pseudo-random number generators. Devices have been constructed that link the output signal of radioactive decay processes or light intensities of so-called ‘lava lamps’ through encryption algorithms and to digit scramblers. To some extent, these approaches ought to be taken with the proverbial pinch of salt since the benefits that can be gained from the additional level of randomisation can hardly be in any proportion to the extreme effort and thus cost involved. This is certainly true for applications of Monte Carlo methods in finance but may be different when it comes to security critical encryption uses of random numbers. This chapter is not meant to be an exhaustive overview over available pseudo-random number generation methods. There are many excellent books on this subject alone [Knu81, PTVF92, Tez95, Nie92], and the reader is referred to them for details of individual number generators. Section 7.1 is mainly for the entertainment of all those of always wondered about the difference between chaos and randomness, but never dared to ask. In section 7.2, a little historical detour to the beginnings of computer generated pseudo-random numbers is taken. Then, in section 7.3, I briefly outline the most basic principle of pseudo-random number generation. Following that, I acknowledge the probably most frequently used number generators around, namely Ran0 to Ran3 as denoted in [PTVF92].
7.1 Chaos The general principle of pseudo-random number generation is as follows. Given the current value of one or more (usually internally stored) state variables, apply a mathematical iteration algorithm to obtain a new set of values for the state variables, and use a specific formula to obtain a new uniform (0, 1) variate from the current values of all the state variables. This kind of process is mathematically also known as a discrete-time dynamical system. A simple example of a one-dimensional discrete dynamical system is the logistic map. It was originally used by P. F. Verhulst in 1845 to model the development of a population in a limited environment [May76], and is known as a consequence of its non-linearity to produce chaotic dynamics for certain choices of parameters. The logistic map gives rise to discrete dynamics by the following algorithm. Given a number xn ∈ (0, 1), we have xn+1 = µxn (1 − xn )
(7.1)
67
7.1 Chaos 1
xn+1
0.75
0.5
0.25
logistic map f(x) = 4 x ( 1 - x ) identity I(x) = x
0 0
0.25
0.5 x
0.75
1
Figure 7.1: The logistic map for µ = 4. for some µ ∈ (0, 4] which is shown in figure 7.1. For any value of µ and x0 , the system (7.1) will converge to a so-called attractive invariant set. This means, skipping some initial transient behaviour, the system (7.1) will only produce iterates that take on values from this invariant set. As long as the invariant set is (finitely) piecewise continuous, we can remap it relatively easily to the interval (0, 1), and thus obtain the desired uniform (0, 1) variates (whereby we have yet not ascertained to what extent these meet the requirement of serial decorrelation etc.). The first question that arises is what value of µ should we use. For this purpose, I plot in figure 7.2 the invariant set for values of µ from 0 to 4. As we can see in the figure, the most promising value for µ is 4 since then the entire interval (0, 1) appears to be filled homogeneously. In fact, the value µ = 4 can be shown to give rise to a strange attractor for the dynamical system 7.1, and to fill the uniform interval (0, 1) with Lebesgue measure 1 [GH83]. Does this mean we can use the logistic map as a pseudo-random number generator? Sadly, no. There are two substantial problems with it and I shall now briefly look at them individually. First, have a look at figure 7.3. What you see are iterates from the logistic map for µ = 4. Admittedly, they look rather random indeed, and it is this similarity between chaos and randomness which gives rise to the common misunderstanding that chaotic systems behave randomly. Now, have a closer look at the ordinate level of 0.75. There is clearly some kind of structure. What happens is that any iterate that comes very close to 0.75 is succeeded by more iterates very nearby. We see from equation (7.1) that x∗ = 3/4 is actually a fixed point under the logistic map for µ = 4. However, this fixed point is not stable under the dynamics of (7.1). This means that any point in an arbitrarily small vicinity of x∗ is gradually repelled from the fixed point, until it eventually starts orbiting all over the unit interval again. As a consequence, a number generator based on the logistic map would never return the value 0.75, but instead, whenever it happens to come close to this value, there will be a number of variates following with values very nearby, displaying a temporary near-periodicity. What’s more, there are
68
Pseudo-random numbers 1
Attractive invariant set
0.75
0.5
0.25
0 0
1
2
3
3.5
3.8
3.95
4
µ
Figure 7.2: The attractive invariant set of the logistic map for values in µ from 0 to 4. Note that the abscissa has been scaled according to µ → − ln(4.25 − µ).
actually uncountably infinitely many other values that are also excluded by the map. Firstly, there are all of the predecessors of 0.75, i.e. all the values that would actually end up right on 0.75 under the iteration rule (7.1). Going back from x∗ = 3/4 one iteration we have { 1/4, 3/4}. One more backwards iteration gives us { 1/2 −
√
3/ , 3/ , 1/ + 4 4 2
√
3/ }. 4
In each iteration that we look backwards out from x∗ ,
we have from then on two new predecessors of the unstable fixed point for each one that we already had. Therefore, there are limi→∞ 2i many points excluded from the attractive invariant set of the logistic map for µ = 4. What we are left with after taking all of those points out of the unit interval is what is known as a Cantor set. It is impossible to map a Cantor set back to a continuous interval with any sensible function that could be used on a computer. Also, the predecessors of x∗ are not the only points that are being avoided by the logistic map. In addition, there are also all the points that represent short periodic orbits (i.e. fixed points of the iterated map), and they also have domains of √ temporary near-periodicity in their vicinity. Take for instance the value x0 = 18 (5 − 5). Its iterate √ √ is x1 = 18 (5 + 5) which in turn leads back to x2 = 18 (5 − 5) = x0 , and thus we have a period-2 orbit. This period-2 orbit is unstable and therefore we have another set of repelling points embedded in the chaotic attractive invariant set for µ = 4. Equally, there are points forming short periodic orbits for many other cycle lengths, and all of them are unstable at µ = 4. This is one of the most defining
69
7.1 Chaos 1
iterate value
0.75
0.5
0.25
0 0
25000
50000 iteration
75000
100000
Figure 7.3: Iterates from the logistic map at µ = 4. points of chaotic motion: embedded inside the attractive invariant set of chaotic dynamics, known as a strange attractor, there is always an infinity of unstable periodic orbits. The second problem with a nonlinear dynamical system such as the logistic map is that of the nonuniformity of the invariant measure. For a random number generator, the invariant measure ψ(x) is simply the probability density that the next draw will be in the interval [x, x+dx], and for a discretely iterated system it is the long-term average of points arriving in that interval. The attentive reader may have noticed that the points in figure 7.3 are a lot denser at the top and at the bottom of the diagram. This is a symptom of the fact that for µ = 4 the invariant measure of the logistic map is given by 1 ψµ=4 (x) = p π x(1 − x)
(7.2)
[vN51, GH83], which is singular at x = 0 and x = 1. Nonetheless, if we transform from x to the new variable α by setting π x = sin α 2 2
⇐⇒
α=
√ 2 arcsin x , π
(7.3)
70
Pseudo-random numbers
we can calculate the density η(α) of the new variable α on (0, 1) according to ψ(x) dx = η(α) dα
(7.4)
dx dα π sin π2 α cos π2 α p η(α) = π x(1 − x) η(α) = ψ(x(α))
η(α) = 1 .
(7.5)
In other words, the transformed variable α is uniformly distributed on (0, 1), and thus we might be able to use it as a generator for uniform (0, 1) pseudo-random numbers. Alternatively, we could use the untransformed variates x and correct the probality density associated with each draw as explained in section 9.2, and thus avoid the evaluation of an inverse trigonometric function1 . However, the issue of embedded unstable periodic orbits and all their pre-iterates, i.e. all the points that would map onto them, cannot be removed in any way that we can implement on computers2 . In the early days of Monte Carlo simulations on electronic computers, these islands of regularity in the stream of generated numbers were not considered a major issue, since they can be shown to be of Lebesgue measure zero, and the very iteration formula (7.1) for µ = 4 was used [UvN47, vN51] as an alternative to the midsquare method mentioned in section 7.2. However, since the mid-1940s, a lot of progress has been made in the area of nonlinear dynamics, and the phenomenon of temporary near-periodic behaviour in the vicinity of unstable periodic orbits (also known as intermittency) is much better understood. For financial simulations, we definitely don’t want to employ a sequence that displays features of such obvious involuntary regularity as is shown in figure 7.3.
7.2 The mid-square method The first ever algorithm for the computer generation of pseudo-random numbers was proposed by John von Neumann, and is known as the mid-square method [Ham51]. The procedure is as follows. Take a uniform (0, 1) number x0 in a four-digit representation. The next variate xn+1 is calculated from xn by taking the square of xn and extracting the middle four digits. In other words, from x0 = 0.9876 we obtain x20 = 0.97 |{z} 5353 76 x1 = 0.5353 x21 = 0.28 |{z} 6546 09 x2 = 0.6546 1
albeit that it may not be worth the trouble, see section 14.6 Mathematically, this may be possible by defining a function that is discontinuous everywhere but such a function cannot easily be implemented on a computer. 2
71
7.3 Congruential generation
and so on. Unfortunately, this procedure is very likely to end up in a short periodic orbit or at 0, depending on the initial value x0 , which was recognised early on [For51]. In fact, for the above starting value it is absorbed at 0 after 50 iterations. Even when a starting value is used that does not enter a short periodic orbit soon, and does not map into 0, this method tends to produce a disproportionate frequency of small numbers [Sob94].
7.3 Congruential generation The most commonly used pseudo-random number generation methods are essentially piecewise linear, or piecewise affine, to be precise. The basic idea is to produce integer values mn on a given interval [0, M − 1], and to return a uniform (0, 1) variate un by rescaling3 . The next integer variate is calculated by mn+1 = (a mn + c) mod M .
(7.6)
Note that equation (7.6) is piecewise affine with the same multiplier over all pieces. Thus, it preserves the volume of any given subinterval of [0, M − 1] which is why it is called a congruential generator. Unlike the nonlinear methods that were presented above, there is no absorption into a fixed point or short periodic orbit for congruential generators. In fact, since all calculations are carried out using integer arithmetics, and since a and M must be chosen to be coprime, the system (7.6) cannot give rise to fixed points at all. The constant a is typically chosen to be sufficiently large in order to have any two close initial values very quickly wrap around the unit interval repeatedly and thus appear to be decorrelated. This in turn means that the iteration map (7.6) consists of many nearly vertical lines. For all the specific examples of the multiplier a discussed below, the iteration map would actually appear to be a completely filled unit square, unless the diagram is enlarged greatly, which is why I omit to show it. Incidentally, the transformation (7.3) of the logistic map (7.1) for µ = 4 results in the iteration scheme ( αn+1 =
2αn
for
2 − 2αn
for
αn < 1 2
1 2
≤ αn ≤ 1
(7.7)
which has the same structural features as αn+1 = (2 αn ) mod 1 .
(7.8)
The small multiplier 2 in this form makes it somewhat clearer that the logistic map displays the feature of comparatively slow separation of two initial values near unstable periodic orbits or near the fixed point at x∗ = 3/4 in figure 7.3. Frequently, the constant c in equation (7.6) is chosen to be zero, whence we commonly encounter the name linear congruential generator. There is a lot of literature on good choices for a and M , and n The common method for rescaling is to set un = m M . Since mn can, however, take on the value 0, which we usually n +1 want to avoid, I recommend to rescale according to un = m M +1 .
3
72
Pseudo-random numbers
not all of it is trustworthy. In IBM’s early computing days, it used to deliver its mainframe systems equipped with its infamous RANDU generator which uses a = 65539, M = 231 and m0 = 1. This generator has meanwhile repeatedly been reported to be highly inadequate [PTVF92, Sob94, FMM77, Tez95]4 . Sadly, it has been copied to a variety of other computer systems, including even the Russian ES system [Sob94]. Since it is well conceivable that other computer manufacturers have slipped up similarly5 , and since there is according to Murphy’s law a tendency for mistakes to proliferate, I advise the reader never to rely on black box number generator that come with any one system and allegedly have been tested. Note that this kind of mistake is extremely easy to make. Again, please don’t trust any number generator that you can’t find a proper reference to, even if it looks sophisticated and similar to the trustworthy ones given in, e.g. [PTVF92]. Not all is doom and gloom, however. A simple choice for the constructing multiplier and modulus that does work is a = 517 , M = 240 , and m0 = 1, and has been used successfully in Russian 40-bit computers [Sob94]. Its period is 238 .
7.4
Ran0 to Ran3
The seminal masterpiece by Press, Teukolsky, Vetterling, and Flannery [PTVF92] is a very good source of reliable number generators that have been well tested and are well understood. The simplest of their suggested number generators, Ran0, is given by equation (7.6) with a = 75 = 16807, c = 0, and M = 231 − 1. This choice of parameters was proposed by Park and Miller [PM88] as a minimal standard generator and goes back to IBM’s GGL generator [Tez95, LGM69] from 1969. There are some technicalities involved with the issue of overcoming the roundoff problem in the multiplication in (7.6) for the given parameters on 32-bit computers. This is typically done by the aid of Schrage’s method, and the reader is referred to [PTVF92] for the details. Ran1 is an enhancement of Ran0 using a careful shuffling algorithm. Note that any kind of enhancement of a pseudo-random number generator has to be done with utmost care, and should be left to number theoreticians. The reason is that any non-linear alteration or modification of the numbers given by one number generator is more than likely to result in disaster6 . The message here is: Kids, don’t do this at home! Sobol’ [Sob94] reports that IBM’s RANDU uses M = 229 , whilst [PTVF92] and [Tez95] report it to use M = 231 . I don’t know which is accurate or whether it makes a difference. I simply advise the reader to avoid any number generator that uses the multiplier a = 65539. 5 [PTVF92] report that one popular 32-bit PC-compatible compiler comes with a severely flawed pseudo-random number generator. Apparently, a reasonable number generator has been used as the basis for the compiler vendor’s own design which was to add a byte swapping mechanism. This kind of action is always a dangerous thing to do and in this case ruined the number generator. 6 You may, if you want, compare it to the problem of interaction of two individually chaotic systems (although randomness is not chaos, as was demonstrated in the previous sections) which in most cases gives rise to stable periodic behaviour. 4
7.5 The Mersenne twister
73
Ran2 is based on the idea of coupling two linear congruential generators, to construct one of a much longer period [L’E88]. And finally, there is Ran3 which is based on a subtractive method suggested by Knuth [Knu81]. I won’t go into details for either of them because I don’t think I could explain them any beter than it is already done in [PTVF92].
7.5 The Mersenne twister Another random number generator technique that has become increasingly popular recently is the Mersenne Twister. The name is to indicate that the period of the sequence is a Mersenne number, i.e. a prime number that can be written like 2n −1 for some n ∈ N, and that it belongs to the class of ‘Twisted Generalized Feedback Shift Register’ sequence generators. The period of the Mersenne Twister as published [MN98] and available [MN97] is 219937 − 1. In order to give you a feeling for this number, imagine that we started at the time of creation of the universe a computer producing 1 billion numbers per second7 from the Mersenne twister sequence. The fraction of the full period that this computer would have produced by now is a decimal number with 5975 digits of zeros behind the decimal point, prior to any non-zero digits. In other words, this computer could continue to draw numbers for many thousand life cycles of your average solar system between its formation and collapse into a black hole before beginning to repeat the sequence. Clearly, for all practical purposes, this number generator can be assumed to have infinite periodicity. The Mersenne twister sequence is guaranteed to have equidistribution properties in at least 623 dimensions. As George Marsaglia put it in 1968, ‘Random numbers fall mainly in the planes’, for all random number sequences there exists an embedding dimensionality in which, in the right projection, all sample points appear to lie in hyperplanes. This can have fatal consequences for a Monte Carlo calculation if the problem that is evaluated just so happens to be susceptible to the used sequence’s regularity. The higher the embedding dimension is which produces those patterns, the safer will be the underlying number generator for general purpose use. I do not intend to go into details about the internal mechanics of this particular generator; suffice it to say that it tries to utilise as much of the existing number theory to produce a reliable number generator. To quote the authors themselves: MT uses many existing ideas. In particular, we are thankful to the following persons. N. Yoneda, P. L’Ecuyer, R. Couture, H. Niederreiter, P. Hellekalek, M. Fushimi, S. Tezuka, Y. Kurita, D. Knuth, H. Leeb, S. Wegenkittl, T. Cooper, M. Rieffel, H. Enomoto, and many many persons who gave us valuable comments, including the users of TT800, a previous twisted generator.
In all of my tests and experiences this sequence engine performed well. Since code is freely available for it [MN97], and since it is no slower than any of the other pseudo-random number generators8 , it is 7
Even though at the time of this writing computers running at 1GHz CPU clock frequency are readily available, software running on such fast hardware could only produce Mersenne twister numbers at a rate of less than one hundred million draws per second. 8 It is in fact faster than most other reliable pseudo-random number generators.
74
Pseudo-random numbers
recommended to integrate it into your general purpose library. It is certainly worth having it, and even if only as a backup should you ever wish to cross-test results against a number generator whose code base was not taken from the seminal reference ‘Numerical Recipes in C’ [PTVF92].
7.6 Which one to use? All of them, or at least more than one of them. Sadly, many people who use Monte Carlo methods underestimate the importance of the underlying number generator greatly, or have some kind of deep trust in whoever built the number generator they are using from within a black box. This is not to say that the sequence generator is the most difficult part of the design of a Monte Carlo procedure, on the contrary. It is, however, one link in a chain of techniques that together comprise a Monte Carlo simulation, and the nature of chains is that they are only as strong as their weakest link. And unlike some add-on parts of the chain such as control variates etc. which are supposed to just add strength but their weak performance does not break the fundamentals of the calculation, the reliability of the number generator is crucial. I hope that by the end of this chapter the reader is sufficiently aware of the basic principles that ensure that, as a matter of fact, all pseudo-random number generators are flawed. How suitable they still are for any one application, depends on the very problem that is being tackled, and it is virtually impossible to foresee where a number generator may break down. Therefore, I advise any user of Monte Carlo methods to have a small library of number generators available, and rather than re-running a calculation with a new seed for any one generator, flick a switch that makes the simulation use a different number construction method altogether. That’s the way to do it.
Chapter 8 Low-discrepancy numbers In the light of the fundamental difficulties involved in the generation of truly random numbers mentioned in chapter 7, we may ask ourselves why do we need randomness? The answer lies with the need of representing multi-dimensional joint distributions. As an easy example, let us consider the incremental path construction of standard Brownian motion W (ti+1 ) = W (ti ) +
p ∆ti zi
(8.1)
with zi ∼ N (0, 1) over the time interval [0, tn ] for i = 0..n − 1. Evidently, any serial correlation between the drawn normal variates zi will give rise to a bias or undesirable regularity in the constructed path. If many paths constructed in this way are used as the basis for the valuation of a path-dependent financial derivative, for instance, any serial correlation is likely to give rise to a mispricing. We can, however, view this kind of problem from a different perspective. What we are really sampling here is a function of a vector argument z whose elements are to be taken from an n-dimensional standard normal distribution. In this formulation, there is absolutely no reason why subsequent draws of n-dimensional vector variates z would have to be serially uncorrelated. The need for perfect decorrelation and thus randomness is merely an artefact of the custom to construct multi-dimensional vector variates from one and the same underlying one-dimensional number generator. For number generation methods that allow for a given dimensionality of the problem at hand, taking previous vector draws into account, and thus making subsequent vector draws serially correlated in order to avoid the inevitable clusters and gaps of (pseudo-)random numbers, can actually aid the equidistribution property of the number sequence generator. This is the essential idea behind the concept of low-discrepancy numbers. Unlike pseudo-random numbers, low-discrepancy numbers aim not to be serially uncorrelated, but instead to take into account which points in the domain to be sampled have already been probed. The mathematical foundations of low-discrepancy sequences go back to H. Weyl in 1917 [Sob94] and many number theoreticians have worked in this field although I name but a few of the better known sequences: Halton, Faure, Haselgrove, Niederreiter, and Sobol’ [Sob67, Sob76, BF88, BFN94, Nie88, Nie96, Nie92, Tez95]. Low-discrepancy numbers have become a popular tool for financial Monte Carlo calculations since the early 1990s [BMW92]. 75
76
Low-discrepancy numbers It has been shown that number sequences can be generated that enable us to do quasi-Monte Carlo
calculations which, given certain smoothness conditions of the function to be integrated, converge not as one over the square root of the number of samples taken, i.e. ∝
√1 , N
but instead much more closely
d
to one over N , namely ∝ c(d) (lnNN ) . This, even for a large dimensionality d, is asymptotically much faster than ∝
√1 . N
The only problem is that the coefficient c(d) can depend on the dimensionality and
thus for any one high-dimensional calculation we cannot know in advance if the use of low-discrepancy numbers will be beneficial with respect to the accuracy required for the specific computation. At this point, number-theoretical results alone no longer tell us if any particular low-discrepancy sequence will give a speedup in convergence and we have to rely on empirical results. In this chapter, I first explain the number-theoretical concept of discrepancy. Then, we introduce the Halton sequence which is probably the easiest low-discrepancy number generation method to describe. Next, I discuss the Sobol’ sequence. Following that, we briefly discuss the Niederreiter (1988) method. For further details and for other low-discrepancy number generation methods the reader is advised to refer to the books by Tezuka [Tez95] and Niederreiter [Nie92]. Then, some empirical evidence will be given that Sobol’ numbers, if adequately initialised, can be used in high dimensions, conversely to common belief. I will also try to explain the origin to what I think is a misunderstanding in the literature that they begin to fail as and when you start using dimensionalities above a few dozens.
8.1 Discrepancy A measure for how inhomogeneous a set of d-dimensional vectors {r i } is distributed in the unit hypercube is the so-called discrepancy. A simple geometrical interpretation of the number-theoretical definition is as follows. Generate a set of N multivariate draws {r i } from a selected uniform number generation method of dimensionality d. All of these N vectors describe the coordinates of points in the d dimensional unit hypercube [0, 1]d . Now, select a sub-hypercube S(y) by choosing a point y delimiting the upper right corner of the hyper-rectangular domain from 0 to y. In other words, the sub-hypercube S can be written as S(y) = [0, y1 ) × · · · × [0, yd ). Next, let nS(y) denote the number of all of those draws that are in S(y), i.e. nS(y) = =
N X
1{ri ∈S(y)}
i=1 N Y d X
1{yk ≥rik } .
(8.2)
i=1 k=1
In the limit of N → ∞, we clearly require perfect homogeneity from the sequence generator which means d
nS(y) Y lim = yi N →∞ N i=1
(8.3)
77
8.2 Halton numbers
for all y ∈ [0, 1]d . The above equation simply results from the fact that for a perfectly homogeneous and uniform distribution on a unit hypercube the probability to be in a subdomain is equal to the volume of that subdomain, and the volume V of S(y) is given by the right hand side of equation (8.3). With nS(y)/ N
these definitions, we can now compare
with V (S(y)) for any one given y. In order to obtain
a measure for the global discrepancy of the number generator, we still need to choose an error norm over all possible y in the unit hypercube. With respect to the L2 -norm, this gives us (d)
d Y
nS(y) − yk N k=1
Z
TN = [0,1]
d
12
!2
dy .
(8.4)
Another frequently used discrepancy measure is the one resulting from the above procedure involving the L∞ -norm: (d) DN
d n S(y) Y − yk = sup y∈[0,1]d N k=1
(8.5) (d)
(d)
Clearly, by the nature of the underlying norms used for the respective definitions of TN and DN , we have (d)
(d)
DN ≥ TN .
(8.6)
For numerical tests, the L∞ discrepancy is rather cumbersome to evaluate. However, as is demonstrated in appendix 8.8.1, the discrepancy with respect to the L2 -norm can be evaluated to the explicit formula
(d)
TN
2
=
N d N d 21−d X Y 1 XY 2 (1 − max(r , r )) − 1 − r + 3−d ik jk ik N 2 i,j=1 k=1 N i=1 k=1
,
(8.7)
where rik is the k-th element of r i . In appendix 8.8.2, I show that the expected squared discrepancy for truly random numbers is h i 1 −d (d) 2 E TN = 2 − 3−d . N
(8.8)
We now arrive at the number-theoretical definition of low-discrepancy sequences. A sequence in [0, 1]d is called a low-discrepancy sequence if for all N > 1 the first N points in the sequence satisfy (d)
DN ≤ c(d)
(ln N )d N
(8.9)
for some constant c(d) that is only a function of d.
8.2 Halton numbers The idea behind Halton numbers is to use the representation of a given generating integer γ in a different number base for each dimension. Of course, the mentioned integer has to be a different one
78
Low-discrepancy numbers
for each new vector draw. An easy natural choice for this constructing integer is simply the number n for the nth draw, γ(n) := n, but any other choice of using a new integer with each draw such as the Gray code G(n) discussed in section 8.3.3 would work, too. In order to prevent any asymptotic pairwise periodicity, the number bases are chosen to be the prime numbers, of which one has to be precalculated for each dimension. The algorithm to construct a new vector draw of Halton numbers is as follows. 1. For each of the required dimensions i = 1..d, find the representation of γ(n) in the associated prime number base pi , i.e. find the coefficients ak in mni X
γ(n) =
aki pk−1 , i
(8.10)
k=1
with all of the aki < pi and mni chosen large enough to make sure that all non-zero digits of γ(n) in the number base pi are accounted for. 2. To construct the coordinate uni for dimension i of the nth uniform vector draw, the sequence of calculated coefficients is now inverted and used as multipliers of fractions in the number base pi , i.e. mni X
uni =
aki p−k . i
(8.11)
k=1
For instance, if we choose the prime numbers 2, 3, 5, 7 as the basis in the four dimensions of a fourdimensional Halton sequence, the 37th draw is constructed as follows for γ(n) = n. base 2 3
n = 3710 in base 1001012 11013
uni 0.1010012 0.10113
= =
5
1225
0.2215
=
7
527
0.257
=
1·2
−1
−2
+0·2
−1
1·3
−3
+0·2
−4
+0·3
+1·3
−3
−1
−2
+1·2
−2
2·5
+2·5
−5
+0·2
−4
+1·3 −3
+1·5
2 · 7−1 + 5 · 7−2
−6
+1·2
=
0.640625
=
0.382716
=
0.488000
=
0.387755
An actual implementation of the algorithm to create the next draw of the Halton sequence is given in code example 8.1. The variable sequenceCounter is the index of the next vector draw, i.e. it represents n in the discussion above. The precalculated prime numbers are stored in the array primeNumbers.
8.3
Sobol’ numbers
The construction of Sobol’ numbers [Sob67] is somewhat more involved. Again, a set of incommensurate basis numbers is used. This time, however, a different kind of multiplication determines the meaning of incommensurate. Whereas for Halton numbers, the basis numbers simply had to be incommensurate with respect to ordinary multiplication, for Sobol’ numbers, the basis numbers are compared with respect to binary multiplication modulo two.
79
8.3 Sobol’ numbers const vector<double>& Halton::nextUniformVector( void ) { unsigned long b, i, k; double f, h; for (++sequenceCounter, i = 0; (i < dimensionality); ++i) { for ( k = sequenceCounter, b = primeNumbers[i], f = 1., h = 0.; f /= b; h += (k%b)*f; } sequenceVector[i] = h; } return sequenceVector; }
(k);
k/=b ) {
Code example 8.1: Code sample for the generation of the next vector draw of the Halton sequence.
8.3.1 Primitive polynomials modulo two The theory of Sobol’ numbers starts with modular integer arithmetic. Two integers i and j are called congruent with respect to the modulus m, i.e. i,j
mod m
(8.12)
if and only if the difference i − j is divisible by m. Clearly, the numbers 0..m − 1 are sufficient to represent the result of any multiplication or addition in the modulus m, due to the congruence relation (8.12). For m being prime, the combination of addition and multiplication modulo m, plus a neutral element with respect to both, is also called a finite commutative ring which is isomorphic to a Galois Field with m elements, GF [m]. A polynomial P (z) of degree g, P (z) =
g X
ak z g−j ,
(8.13)
j=0
is considered to be an element of the ring GF [m, z] of polynomials over the finite field GF [m] if we assume all of the coefficients ak to be ∈ GF [m]. In other words, all algebra on the coefficients ak is to be carried out modulo m. This means, for instance, that (z + 1)(z + 1)(z 2 + z + 1) , z 4 + z 3 + z + 1 mod 2 .
(8.14)
A polynomial P (z) of positive degree is considered to be irreducible modulo m if there are no other two polynomials Q(z) and R(z) which are not constant or equal to P (z) itself such that P (z) , Q(z)R(z)
mod m .
(8.15)
An irreducible polynomial modulo m in GF [m, z] is the equivalent to a prime number in the set of integers. The order of a polynomial P (z) modulo m is given by the smallest positive integer q for which P (z) divides z q − 1, i.e. q = inf0 q
n
0 q 0 z q − 1 , P (z)R(z)
mod m
o
(8.16)
80
Low-discrepancy numbers
for some non-constant polynomial R(z). An irreducible polynomial P (z) of degree g is also considered to be primitive modulo m if its order is mg − 1. Note that not all irreducible polynomials are also primitive, although (especially for m = 2) most of them are. The importance of primitive polynomials modulo two is given by two separate facts. Firstly, algebraic manipulations modulo two, i.e. binary algebra, is particularly well suited to implementation on today’s digital computers. Secondly, for any primitive polynomial of degree g there are recurrence relations for obtaining a new random bit from g preceding ones [PTVF92]. In other words, the use of a distinct primitive polynomial modulo two for each of the required dimensions of the vector of uniform numbers, makes it possible to generate a sequence of vectors in which the sampling happens both uniformly over all of the dimensions, but in each dimension, we also have uniform use of all of the binary digits. For further information on the distinction between irreducibility and primitivity, see, e.g. [Tez95, Chi00]. The calculation of primitive polynomials can be rather involved. Whilst there are limited tables of primitive polynomials available [Wat62, PTVF92, J¨ac97], I provide a list of all primitive polynomials modulo two up to degree 27 on the accompanying CD. This amounts to a total number of 8 129 334 primitive polynomials which should be more than enough for all practical applications.
8.3.2
The construction of Sobol’ numbers
The generation of Sobol’ numbers is initially carried out on a set of integers in the interval from 1 to a power of two minus one, say [1, 2b − 1]. As you may imagine, b simply represents the number of bits in an unsigned integer on the given computer and should typically be 32, which amounts to the set of attainable integers being given by all those in the range [1, 4 294 967 295]. We will denote the nth draw of one such Sobol’ integer in dimension k as xnk . The final conversion to a uniform variate ynk ∈ (0, 1) is done by dividing xnk by 2b as a floating point operation, i.e. xnk ynk := b , ynk ∈ (0, 1) , xnk ∈ Z[1, 2b − 1] (8.17) 2 By construction, the only Sobol’ variate that could ever be exactly zero1 is the zeroth draw (more on the meaning of this later), and this holds for all dimensions. Therefore, we can explicitly exclude the possibility that any one of the drawn integers is actually zero by simply skipping the zeroth draw. For each of the d dimensions, the basis of the number generation is given by a set of so-called direction integers, of which there is one for each of the b bits in the binary integer representation. It is conducive for the following to view all of the direction integers as b-wide bit fields. Let us denote the lth direction integer for dimension k as vkl . Additional constraints on the bit field representing vkl are that only the l leftmost2 bits can be non-zero, and that the lth leftmost bit of vkl must be set. The actual 1
This is to mean for both the original Sobol’ algorithm as well as the Antonov-Saleev modification using the conventional Gray code, but also for all other methods that choose the generating integer γ(n) such that γ(0) = 0. 2 The leftmost bits in a bit field representing an integer are the most significant ones. On standard contemporary computers, the number of bits in an integer is 32. Thus, an unsigned integer with only the leftmost bit set would correspend to the number 231 = 2147483648, and all bits being 1 corresponds to 4294967295.
81
8.3 Sobol’ numbers
number draws will later on be calculated by binary addition modulo two of some of these direction integers, which makes it clear that each vkl can only affect the l leftmost bits in the drawn integer xnk , and that it definitely influences the lth leftmost bit of xnk . Binary addition of integers modulo two, which amounts to bitwise addition without carry, is a particularly fast operation on contemporary computers known as Exclusive OR, and is usually abbreviated as XOR. The key to the generation of Sobol’ numbers is the calculation of the direction integers. This involves the binary coefficients of a selected primitive polynomial modulo two for each dimension. Let the primitive polynomial modulo two for dimension k be pk . Denote the degree of this polynomial as gk . Let the coefficient of the highest monomial in pk be ak0 , and so forth down to ak gk , i.e. pk (z) =
gk X
akj z gk −j .
(8.18)
j=0
Note that the fact that ak0 ≡ 1 is a simple consequence of pk (z) being of degree gk . Now we come to the setting up of the direction integers. In each dimension with its associated primitive polynomial pk , the first gk direction integers vkl for l = 1..gk can be chosen freely, within the above mentioned two constraints. All subsequent ones are constructed from the following recurrence relation: gk ⊕2 X vk (l−gk ) vkl = ⊕2 akj vk (l−j) 2gk j=1
for
l > gk
Hereby, the operator ⊕2 stands for the XOR operation3 , and the notation
(8.19)
... ⊕2 P
indicates a whole se-
...
quence of XOR operations, or binary additions without carry, in analogy to the conventional sum P operator . In other words, the direction integer vk (l−gk ) is right-shifted by gk bits, and then XORed with a selection of the (unshifted) direction integers vk (l−j) for j = 1..gk (controlled by which of the coefficients akj are set), to obtain vkl . Note that the highest order coefficient ak0 is not actually used in the recurrence relation (8.19), and that since the lowest order coefficient ak gk is always set, the direction integer vk (l−gk ) will always enter vkl . This is the reason why the highest and lowest coefficient of the polynomial pk is usually not included in its encoding, provided that its degree is known [PTVF92]. If you made it this far, and managed to follow the above explanations, you’ll be relieved to see how simple the actual construction of the Sobol’ integers xnk turns out, given the above preliminaries. Just like for the construction of Halton numbers, we need a new unique generating integer γ(n) for each new draw. An easy choice of such an integer for the nth draw is n itself, i.e. γ(n) := n which amounts to the original algorithm published by Sobol’ [Sob67]. However, any other method of ensuring a new integer for each new draw such as the Gray code, i.e. γ(n) := G(n), is equally possible. Given the generating integer of the nth draw, the Sobol’ integers for all of the d dimensions are given by xnk :=
d X
⊕2
vkj 1{j th bit (counting from the right) of γ(n) is set} .
j=1 3
Just in case you don’t know this already but still care: The “⊕2 ” operation’s equivalent in C is “ˆ”.
(8.20)
82
Low-discrepancy numbers
In other words, depending on which of the bits in the binary representation of γ(n) are set, the direction integers are simply XORed, to produce the Sobol’ integer xnk . The final transformation to a uniform floating point number in the interval was already given by the simple division in equation (8.17). It may be clear from formula (8.20) that we need to ensure γ(n) 6= 0
(8.21)
in order to prevent any of the ynk being exactly zero. By the nature of the construction algorithm of the direction integers, no other value for γ(n) can result in any xnk and thus ynk being zero, whence condition (8.21) is sufficient to ensure ynk 6= 0.
8.3.3
The Gray code
Antonov and Saleev contributed to Sobol’ numbers as we know them today by realising that instead of using the binary representation of the sequence counter n directly, any other unique representation of the sequence counter n can be used, too [AS79]. In particular, a bitwise representation of n which switches only one single bit for every increment in n, means that only one single XOR operation is to be carried out for the generation of every integer representing the uniform coordinate of the next vector draw. This kind of encoding of integers is known as a Gray code G(n), named after the engineer Frank Gray who patented this method for the use with shaft encoders in the 1950s. Gray codes are still used with shaft encoders today, and in many communication applications. They are useful wherever a set of parallel electrical wires are used to indicate a number by the individual voltage state of each line. Frequently, such lines are used to transmit the current state of a counter variable, which would only ever increase by one. In such an application, using any encoding whereby more than one bit can change from one number to the next, the tiniest mistiming in the transition of the high-low states from one number to the next will cause the recipient of the signal not to receive a clean increase, but a rapid and spurious sequence of intermediate numbers. As an example, think of the transition from 310 = 0112 to 410 = 1002 . Due to the inevitably limited accuracy of mechanical or electronical components, the receiver is likely to perceive one of the possible sequences from 3 to 4 given in table 8.1. Clearly, none 310 = 0112
210 = 0102
010 = 0002
410 = 1002
310 = 0112
210 = 0102
610 = 1102
410 = 1002
310 = 0112
110 = 0012
010 = 0002
410 = 1002
310 = 0112
110 = 0012
510 = 1012
410 = 1002
310 = 0112
710 = 1112
610 = 1102
410 = 1002
310 = 0112
710 = 1112
510 = 1012
410 = 1002
Table 8.1: Possible transitions from 3 to 4 in standard binary representation. of them are desirable. Using a Gray code in the representation of the integers surmounts this problem because exactly one bit changes in any one increase. It turns out that there is no single unique Gray code. The most commonly used choice for the Gray code is G(n) = n ⊕2 [n/2] .
(8.22)
83
8.3 Sobol’ numbers
In table 8.2, the Gray code of the integers 1 to 7 is given as an example. An interesting feature of the n
n in binary
[n/2] in binary
G(n) in binary
G(n) in decimal
0
000
000
000
0
1
001
000
001
1
2
010
001
011
3
3
011
001
010
2
4
100
010
110
6
5
101
010
111
7
6
110
011
101
5
7
111
011
100
4
Table 8.2: Gray codes. Gray code is that the single bit that changes from G(n) to G(n + 1) is always the rightmost zero bit of the binary representation of n itself. This can be readily verified analytically and is also easy to see in table 8.2. As for the generation of Sobol’ numbers, they are clearly aided by the use of G(n) instead of n as the constructing integer of the nth vector draw. Imagine we have already generated all of the vector draws out to number n − 1, and we have kept in memory the uniform integers x(n−1)· for all of the required dimensions. Since the Gray code G(n) differs from that of the preceding one G(n − 1) by just a single, say the j th , bit (which is the rightmost zero bit of n − 1), all that needs to be done is a single XOR operation for each dimension in order to propagate all of the x(n−1) to xn , i.e. xnk = x(n−1)k ⊕2 vjk .
8.3.4
(8.23)
The initialisation of Sobol’ numbers
The attentive reader may have noticed that there is yet some freedom left in the construction of Sobol’ numbers, namely the specific choice of the free direction numbers. As we recall from section 8.3.2, given the primitive polynomial pk of degree gk associated with the k th dimension, the first gk direction integers can be chosen freely within certain constraints. All of the remaining direction integers are then determined by the recurrence equation (8.19). Since the first gk direction integers thus initialise the entire construction of the sequence, I also call them initialisation numbers. The constraints on the lth initialisation number vkl of dimension k are that only the l leftmost bits can be non-zero, and that the lth bit from the left hand side of the b-wide bit field representing vkl must be 1. Arguably the easiest choice for the initialisation numbers is thus to just have the lth leftmost bit set, and all other bits to be zero, which amounts to what I call unit initialisation, i.e. vkl = 2b−l .
(8.24)
The impact of the initialisation numbers on the homogeneity properties of the entire sequence is not to be underestimated. In 1976, Sobol’ published algebraic conditions that link specific choices of
84
Low-discrepancy numbers
initialisation numbers to certain uniformity properties [Sob76]. A low-discrepancy sequence is said to satisfy property A if for any binary segment (not an arbitrary subset) of the d-dimensional sequence of length 2d there is exactly one draw in each of the 2d hypercubes that result from subdividing the unit hypercube along each of its unit length extensions into half. In other words, assume that in each dimension we divide the interval [0, 1) into the two sub-intervals [0, 1/2) and [ 1/2, 1). This will result in a subdivision of the d-dimensional unit hypercube [0, 1)d into 2d sub-hypercubes. Given any sequential section S of length 2d of the low-discrepancy sequence of d-dimensional uniform variates that starts at some index m = l2d for some integer l, i.e. Sl := (ul2d , ul2d +1 , . . . , u(l+1)2d −1 ), there must be exactly one u ∈ Sl that lies in each and every one of the sub-hypercubes of volume 2−d for the sequence generator to satisfy property A. Property A’ is similar in its definition in that it refers to sections of length 4d being required to provide that a single element is contained in each and every sub-hypercube resulting from the subdivision of the unit hypercube into four equal intervals in each dimension. Of course, if we explicitly exclude the point at the origin u ≡ 0 (as we would in order to prevent a mapping to infinity when transforming to Gaussian variates), the segment of the sequence starting with the first element will have no point in the sub-hypercube [0, 1/2)d and will be most uniform for a length of 2d − 1. The algebraic equations that guarantee properties A and A’ can be solved numerically, and there are precalculated tables of initialisation numbers that provide properties A and A’ up to certain dimensionalities in the literature. The original article [Sob76], for instance, provides initialisation numbers up to dimension 16 for property A and up to dimension 6 for property A’, and some authors have gone through some considerable effort for higher dimensions, namely Paskov and Traub calculated them up until dimension 360 [PT95]. In finance, however, we often face problems that are of very high dimensionality d. In order to benefit from property A, we would need to carry out a Monte Carlo simulation over a number of iterations that is to be of the order of magnitude of 2d . The pricing of an Asian option with daily monitoring and one year to maturity, for instance, represents a Monte Carlo integration in 250 dimensions, one for each trading day in the year. Since 2250 ' 1075 , we would have to iterate as many times as is currently estimated to be the total number of particles in the universe, which is clearly excessive, before we would benefit from property A, not to mention property A’. However, as we will see in sections 8.5 and 8.6 this is not to say that for high-dimensional problems every set of initialisation numbers will work as well as any other. There is a clear benefit from the choice of initialisation numbers that enables the low-discrepancy sequence to start exploring the volume of the unit hypercube early on, rather than initially just focussing on certain areas. I therefore recommend to use initialisation numbers that provide properties A and A’ for the lowest dimensions, and for the higher dimensions, at least to ensure that any regularity in the initialisation set is broken up. One choice of initialisation numbers that does the complete opposite is the aforementioned unit initialisation. Although strictly speaking a valid choice of initialisation numbers, unit initialisation leads to surprisingly bad results for Sobol’ numbers, and should be avoided. A very simple way to generate initialisation numbers that break the regularity, is to use a separate pseudo-random number generator to draw uniform variates ∈ (0, 1) from, and to initialise as follows. Draw u∗kl from a separate uniform
85
8.3 Sobol’ numbers k
gk
1
0
ak0 . . . akg
1
1 · 231
1 · 230
1 · 229
1 · 228
1 · 227
1 · 226
1 · 225
1 · 224
1 · 223
1 · 222
2
1
11
1 · 231
3 · 230
5 · 229
15 · 228
17 · 227
51 · 226
85 · 225
255 · 224
257 · 223
771 · 222
3
2
111
1 · 231
1 · 230
7 · 229
11 · 228
13 · 227
61 · 226
67 · 225
79 · 224
465 · 223
721 · 222
4 5
3 3
1011 1101
1 · 231 1 · 231
3 · 230 1 · 230
7 · 229 5 · 229
5 · 228 3 · 228
7 · 227 15 · 227
43 · 226 51 · 226
49 · 225 125 · 225
147 · 224 141 · 224
439 · 223 177 · 223
1013 · 222 759 · 222
6 7
4 4
10011 11001
1 · 231 1 · 231
3 · 230 1 · 230
1 · 229 3 · 229
1 · 228 7 · 228
9 · 227 31 · 227
59 · 226 47 · 226
25 · 225 109 · 225
89 · 224 173 · 224
321 · 223 181 · 223
835 · 222 949 · 222
8 9 10 11 12 13
5 5 5 5 5 5
100101 101001 101111 110111 111011 111101
1 · 231 1 · 231 1 · 231 1 · 231 1 · 231 1 · 231
3 · 230 3 · 230 1 · 230 1 · 230 3 · 230 3 · 230
3 · 229 7 · 229 5 · 229 7 · 229 7 · 229 5 · 229
9 · 228 7 · 228 11 · 228 3 · 228 13 · 228 1 · 228
9 · 227 21 · 227 27 · 227 29 · 227 3 · 227 15 · 227
57 · 226 61 · 226 53 · 226 51 · 226 35 · 226 19 · 226
43 · 225 55 · 225 69 · 225 47 · 225 89 · 225 113 · 225
43 · 224 19 · 224 25 · 224 97 · 224 9 · 224 115 · 224
225 · 223 59 · 223 103 · 223 233 · 223 235 · 223 411 · 223
113 · 222 761 · 222 615 · 222 39 · 222 929 · 222 157 · 222
14 15 16 17 18 19
6 6 6 6 6 6
1000011 1011011 1100001 1100111 1101101 1110011
1 · 231 1 · 231 1 · 231 1 · 231 1 · 231 1 · 231
1 · 230 1 · 230 3 · 230 1 · 230 1 · 230 3 · 230
1 · 229 3 · 229 3 · 229 7 · 229 1 · 229 5 · 229
9 · 228 13 · 228 5 · 228 13 · 228 3 · 228 11 · 228
23 · 227 11 · 227 19 · 227 25 · 227 13 · 227 7 · 227
37 · 226 7 · 226 33 · 226 5 · 226 39 · 226 11 · 226
97 · 225 37 · 225 3 · 225 27 · 225 7 · 225 43 · 225
97 · 224 101 · 224 197 · 224 71 · 224 23 · 224 25 · 224
353 · 223 463 · 223 329 · 223 377 · 223 391 · 223 187 · 223
169 · 222 657 · 222 983 · 222 719 · 222 389 · 222 825 · 222
20 21 22 23 24 25 26 27 28 29 30 31 32
7 7 7 7 7 7 7 7 7 7 7 7 7
10000011 10001001 10001111 10010001 10011101 10100111 10101011 10111001 10111111 11000001 11001011 11010011 11010101
1 · 231 1 · 231 1 · 231 1 · 231 1 · 231 1 · 231 1 · 231 1 · 231 1 · 231 1 · 231 1 · 231 1 · 231 1 · 231
3 · 230 3 · 230 3 · 230 3 · 230 1 · 230 3 · 230 1 · 230 1 · 230 1 · 230 3 · 230 1 · 230 3 · 230 1 · 230
1 · 229 1 · 229 3 · 229 7 · 229 3 · 229 1 · 229 5 · 229 3 · 229 7 · 229 5 · 229 1 · 229 5 · 229 7 · 229
7 · 228 15 · 228 3 · 228 9 · 228 15 · 228 9 · 228 5 · 228 1 · 228 7 · 228 5 · 228 15 · 228 15 · 228 11 · 228
3 · 227 17 · 227 25 · 227 31 · 227 29 · 227 5 · 227 1 · 227 23 · 227 19 · 227 21 · 227 5 · 227 17 · 227 13 · 227
23 · 226 63 · 226 17 · 226 29 · 226 15 · 226 21 · 226 27 · 226 13 · 226 25 · 226 9 · 226 49 · 226 19 · 226 29 · 226
79 · 225 13 · 225 115 · 225 17 · 225 41 · 225 119 · 225 33 · 225 75 · 225 105 · 225 7 · 225 59 · 225 21 · 225 3 · 225
65 · 224 113 · 224 17 · 224 121 · 224 249 · 224 53 · 224 253 · 224 29 · 224 173 · 224 143 · 224 71 · 224 227 · 224 15 · 224
451 · 223 147 · 223 179 · 223 363 · 223 201 · 223 319 · 223 341 · 223 181 · 223 509 · 223 157 · 223 31 · 223 413 · 223 279 · 223
321 · 222 881 · 222 883 · 222 783 · 222 923 · 222 693 · 222 385 · 222 895 · 222 75 · 222 959 · 222 111 · 222 727 · 222 17 · 222
k
vkl for l = 1..10
Table 8.3: Initialisation numbers providing property A up to dimension 32. random number generator such that wkl := int u∗kl · 2l−1
(8.25)
is odd (simply keep drawing until the condition is met), and set vkl := wkl · 2b−l
for l = 1..gk .
(8.26)
To finish this section I give in table 8.3 a list of initialisation numbers that have been tested to give property A up to dimension 32 and are guaranteed to provide property A’ up to dimension 6 [Sob76]. The columns are as follows: k is the dimension index, gk is the degree of the associated polynomial, ak0 . . . akgk are the coefficients of the polynomial as in equation (8.18), and vkl are the direction numbers. The freely chosen direction numbers, i.e. the initialisation set are in normal black and those derived from the recurrence relation (8.19) are shaded. Note that the polynomial associated with dimension 1 is not strictly a primitive polynomial, similar to the number 1 not strictly being a prime number. This is also the reason why the direction numbers for dimension 1 are not given by the recurrence relation (8.19) but are simply v1l = 2b−l . Further details can be found in the original literature [Sob67]. As for readily available code to construct Sobol’, there is of course the algorithm in [PTVF92]. Using that source code, and the table of primitive polynomials modulo two on the accompanying CD, and the initialisation method described above for the free direction numbers, it shouldn’t be too difficult to create your own high-dimensional Sobol’ number generator. Also, there is a commercial library module available from an organisation called BRODA [KS] that can generate Sobol’ sequences in up
86
Low-discrepancy numbers
to 370 dimensions. In a way, this module can claim to be a genuine Sobol’ number generator since Professor Sobol’ himself is behind the initialisation numbers that drive the sequence, and he is also linked to the company distributing the library. Just before you get any wrong ideas: I am not affiliated with BRODA in any way whatsoever.
8.4
Niederreiter (1988) numbers
H. Niederreiter devised a general framework for number-theoretically constructed sequences of lowdiscrepancy [Nie92]. This contributed greatly to the analysis of the internal mechanisms and helped to understand the similarities and differences of the various number sequences. He also devised several sequence generation algorithms [Nie88, Nie96] of which only one has been implemented [BFN94]. We refer to these numbers as the Niederreiter (1988) sequence. They are from a constructional point of view not too different from Sobol’ numbers. They, too, are based on polynomial arithmetic modulo some base m, and the most frequently used base happens to be 2, not least due to the enormous speed and ease to carry out binary calculations on a modern computer. However, Niederreiter (1988) numbers employ irreducible rather than primitive polynomials4 . Despite the fact that they are theoretically supposed to be superior to Sobol’ numbers in the limit, from an empirical point of view where the start-up rather than the asymptotic performance5 is more relevant, Niederreiter (1988) numbers, in my experience, do not provide quite the same reliability in terms of rapid convergence for high dimensionalities as Sobol’ numbers do.
8.5 Pairwise projections The aim of low-discrepancy number generation methods is to provide a source of vector coordinates that covers a given domain a homogeneously and uniformly as possible. The more homogeneous the underlying number generator, the more accurate and rapidly converging will be a Monte Carlo calculation based on it6 . It has been documented in the literature, that low-discrepancy number generators tend to lose their quality of homogeneous coverage as the dimensionality increases. A particularly striking way to demonstrate this is to plot the projection of a given number of vector coordinates drawn from a number generator onto a two-dimensional projection of adjacent dimensions. In figure 8.1, we show 4
Clearly, since all primitive polynomials are also irreducible, one can just use primitive polynomials for the construction of Niederreiter (1988) numbers, too. This does not, however, remedy the not-so-good performance of Niederreiter (1988) numbers in high-dimensional applications. 5 For practical applications, we are more interested in the realised convergence over the first ten thousand or even 500 million draws, and the question as to which number generator will provide a higher rate of convergence in the asymptotic limit of actually infinitely many draws is not really of great importance. 6 A more mathematical form of this statement is known as the Koksma-Hlawka theorem, but since this relationship is sufficiently plausible by sheer common sense, I won’t go into the details here.
87
8.5 Pairwise projections
the projection of the first 2047 vector draws of various number generators on several two-dimensional uniform intervals. In the first row, we have vector draws from the Mersenne twister as an example
1
8
29
Mersenne twister
94
Mersenne twister
63
Mersenne twister
30
Mersenne twister
9
2
Mersenne twister
62
93
94
63
30
9
2
Sobol’ (regularity breaking initialisation) Sobol’ (regularity breaking initialisation) Sobol’ (regularity breaking initialisation) Sobol’ (regularity breaking initialisation) Sobol’ (regularity breaking initialisation)
93
Sobol’ (unit initialisation)
9
94
62
Sobol’ (unit initialisation)
63
29
Sobol’ (unit initialisation)
30
8
Sobol’ (unit initialisation)
2
1
Sobol’ (unit initialisation)
93
Niederreiter (1988)
9
94
62
Niederreiter (1988)
63
29
Niederreiter (1988)
30
8
Niederreiter (1988)
2
1
Niederreiter (1988)
93
Halton
9 1
8
29
94
62
Halton
63
29
Halton
30
8
Halton
2
1
Halton
62
93
Figure 8.1: Two-dimensional projections of various number generators. for a random pattern for reference. Rows 2 to 5 give the projections of the Sobol’ generator (regularity breaking initialisation), Sobol’ (unit initialisation), Niederreiter, and Halton. In column 1, the projection onto the first two dimensions is shown, followed by dimension 8 versus 9, 29 and 30, 62 and 63, and 93 against 94. The particular dimensions shown were selected randomly. These projections are a first indicator that not all low-discrepancy number generators are reliable when fairly high dimensions are required. For the last three low-discrepancy number generators there is clearly a trend towards
88
Low-discrepancy numbers
clusters and gaps as the dimension increases. Sobol’ numbers with regularity breaking initialisation, however, do not seem to suffer from this problem. There have been reports in various publications that Sobol’ numbers, too, suffer the problem of rapid breakdown of homogeneity in higher dimensions. I believe that this is due to an unfortunate choice of the initialisation numbers used for their construction. In fact, I. Sobol’ himself has been aware of the importance of careful initialisation at least since the early 1970s [Sob76]. Alas, when the financial sector started using low-discrepancy numbers, this wasn’t picked up, whence there is the notion that no low-discrepancy number generator is suitable for high dimensions.
8.6 Empirical discrepancies A more thorough measure for the homogeneity properties of a low-discrepancy number generator than visual inspection of projections is the discrepancy as defined in equations (8.4) or (8.5) in section 8.1. In order to provide even harder evidence that suitably initialised Sobol’ numbers are reliable indeed, even in significantly high-dimensional applications, I show in figures 8.2 to 8.9 the discrepancy with (d)
respect to the L2 -norm TN as defined in equation (8.4) for d = 2, 3, 5, 10, 15, 30, 50, 100 for various number Note that the line denoted as “expectation for truly random numbers” is actually s generators. 2 (d) E TN , as calculated in appendix 8.8.2.
0.01
0.001
Halton
0.0001
1e-05 1024
2 dimensions RAN1 RAN2 RAN3 Mersenne twister Expectation for truly random numbers Halton Niederreiter (1988) Sobol’ (no initialisation liberty for d=2) 2048
4096 (2)
8192
16384
Figure 8.2: TN as defined in equation (8.4).
32768
65536
89
8.6 Empirical discrepancies 0.01
0.001
Halton
0.0001
1024
3 dimensions RAN1 RAN2 RAN3 Mersenne twister Expectation for truly random numbers Halton Sobol’ (unit initialisation) Niederreiter (1988) Sobol’ (regularity breaking initialisation) 2048
4096
8192
16384
32768
65536
32768
65536
(3)
Figure 8.3: TN as defined in equation (8.4).
0.001
0.0001
1024
Halton
5 dimensions RAN1 RAN2 RAN3 Mersenne twister Expectation for truly random numbers Halton Sobol’ (unit initialisation) Niederreiter (1988) Sobol’ (regularity breaking initialisation) 2048
4096 (5)
8192
16384
Figure 8.4: TN as defined in equation (8.4).
90
Low-discrepancy numbers 0.001
Halton
0.0001
1024
10 dimensions RAN1 RAN2 RAN3 Mersenne twister Expectation for truly random numbers Halton Sobol’ (unit initialisation) Niederreiter (1988) Sobol’ (regularity breaking initialisation) 2048
4096 (10)
Figure 8.5: TN
8192
16384
32768
65536
32768
65536
as defined in equation (8.4).
0.0001
Halton
15 dimensions RAN1 RAN2 RAN3 Mersenne twister Expectation for truly random numbers Halton Sobol’ (unit initialisation) Niederreiter (1988) Sobol’ (regularity breaking initialisation) 1024
2048
4096 (15)
Figure 8.6: TN
8192
16384
as defined in equation (8.4).
91
8.6 Empirical discrepancies 0.0001
1e-05
30 dimensions RAN1 RAN2 RAN3 Mersenne twister Expectation for truly random numbers Halton Sobol’ (unit initialisation) Niederreiter (1988) Sobol’ (regularity breaking initialisation)
Halton
Sobol’ (unit initialisation)
1e-06
Sobol’ (regularity breaking initialisation) 1e-07 1024
2048
4096 (30)
Figure 8.7: TN
8192
16384
32768
65536
as defined in equation (8.4).
0.0001 Halton Niederreiter (1988)
1e-05
1e-06
1e-07
50 dimensions RAN1 RAN2 RAN3 Mersenne twister Expectation for truly random numbers Halton Sobol’ (unit initialisation) Niederreiter (1988) Sobol’ (regularity breaking initialisation)
Sobol’ (unit initialisation)
1e-08
1e-09
1e-10 1024
Sobol’ (regularity breaking initialisation) 2048
4096 (50)
Figure 8.8: TN
8192
16384
as defined in equation (8.4).
32768
65536
92
Low-discrepancy numbers 0.0001
Halton Niederreiter (1988)
1e-06
1e-08
1e-10
1e-12
100 dimensions RAN1 RAN2 RAN3 Mersenne twister Expectation for truly random numbers Halton Sobol’ (unit initialisation) Niederreiter (1988) Sobol’ (regularity breaking initialisation)
Sobol’ (unit initialisation)
1e-14
1e-16
Sobol’ (regularity breaking initialisation)
1e-18 1024
2048
4096 (100)
Figure 8.9: TN
8192
16384
32768
65536
as defined in equation (8.4).
It clearly can be seen in figures 8.2 to 8.9 that for low dimensionalities, all of the tested lowdiscrepancy number generators are considerably superior to pseudo-random numbers. However, as the dimensionality increases, this advantage decreases, until around d = 15 the Halton method, the Niederreiter sequence, and Sobol’ with unit initialisation, all appear to be significantly inferior to pseudo-random number methods. This underperformance becomes so dramatic for d = 100 that there are more than ten decimal orders of magnitude between the aforementioned three number methods and plain pseudo-random methods. However, for suitably initialised Sobol’ numbers, there is no deterioration to the extent that they appear inferior to pseudo-random number generator. It is true that the total discrepancy over all of the equally weighted 100 dimensions as shown in figure 8.9 makes well initialised Sobol’ numbers appear as if it is hardly worth bothering. At this point, however, we should bear in mind that in most applications in finance, we are dealing with problems that have a natural ordering in the importance of all of the involved dimensions. A very good example of a simulation problem that decomposes into dimensions of strongly varying importance is that of paths describing Brownian motion. Clearly, if the terminal value of the Brownian motion determines the pay-off of a derivative contract, it is of particular importance in the valuation problem. How the specific features of Brownian motion can be exploited for optimally importance-ranked path construction is discussed in section 10.8 and examples for the good convergence of Sobol’ numbers7 are given in section 10.8.4. The lesson to learn with respect to well initialised Sobol’ numbers is that they will provide a substantial performance boost in the lower dimensions, and will still work at least as well as pseudo7
well-initialised ones, of course
93
8.7 The number of iterations
random number generators in higher dimensions. Thus, we should always try to design the Monte Carlo solution approach such that we can exploit the better convergence in the lower dimensions by assigning them to the problem dimensions of most importance.
8.7
The number of iterations
As we know, for pseudo-random generators, the number of iterations only affects the expected variance of the result by the central limit theorem, i.e. the more the better. For low-discrepancy numbers, the situation is different. Sobol’ numbers, and other number generators based on integer arithmetics modulo two, by construction provide additional equidistribution properties whenever the number of iterations is N = 2n − 1 for some positive integer n. This is easy to see on the unit interval in one dimension where such a choice of draws always results in a perfectly regular distribution of points, and can also be confirmed in the empirical discrepancy diagrams of section 8.6 up to dimension 5.
8.8 Appendix 8.8.1
Explicit formula for the L2 -norm discrepancy on the unit hypercube
Substituting the formula (8.2) for nS(y) into the squared right hand side of equation (8.4), we obtain (d) 2 TN
Z N 1 X = N 2 i,j=1
y∈[0,1]d
d Y
Z N 1 X −2 N i=1
d Y
y∈[0,1]d
N d 1 XY = N 2 i,j=1 k=1
1{yk ≥rik } 1{yk ≥rjk } dy
(8.27)
k=1
Z1
k=1
Z
d Y
y∈[0,1]d
k=1
1{yk ≥rik } yk dy +
yk2 dy
1
1
N d Z d Z Y 1 XY dyk − 2 yk dyk + yk2 dyk N i=1 k=1 k=1
max(rik ,rjk )
rik
(8.28)
0
N d N d 1 XY 1 XY 1 2 = (1 − max(rik , rjk )) − 2 (1 − rik ) + 3−d , 2 N i,j=1 k=1 N i=1 k=1 2
(8.29)
which is identical to equation (8.7).
94
8.8.2
Low-discrepancy numbers
Expected L2 -norm discrepancy of truly random numbers
In order to derive equation (8.8), the expectation of (8.7) for truly random numbers is calculated below. " d # " d # N N h i 1−d X X Y Y 2 2 1 (d) 2 (1 − max(rik , rjk )) − 1 − rik + 3−d (8.30) E TN = E E N 2 i,j=1 k=1 N i=1 k=1 The expectations over products of terms involving random numbers can be replaced by products over expectations when the random numbers are independent. For this to hold, the first sum has to be split into the terms when i = j and when i 6= j. This gives ( d h i Y 1 (d) 2 E TN = N E{xk ∈[0,1]} [1 − xk ] N2 k=1 +N (N − 1)
d Y
(8.31) )
E{xk ,yk ∈[0,1]} [1 − max (xk , yk )]
k=1 d 21−d Y − N E{xk ∈[0,1]} 1 − x2k + 3−d N k=1
1 = N
(
Z −d 2 + (N − 1) 2
1
−2
Z
d
1
(1 − x) dx
d )
1
(1 − x) dxdy
y=0
1−d
Z
(8.32)
x=y
+ 3−d
x=0
" #d Z 1 2 1 1 x = 2−d + (N − 1) 1 − 2 dy N 2 x=y y=0 1−d
−2
1 = N
(8.33)
d 2 + 3−d 3
(
d ) Z 1 2−d + (N − 1) 1 − 1 + y 2 dy − 3−d
(8.34)
y=0
=
1 −d 2 + (N − 1)3−d − 3−d N
(8.35)
=
1 −d 2 − 3−d N
(8.36)
.
Chapter 9 Non-uniform variates Number generators tend to produce uniform variates on the unit interval. Whenever we wish to carry out a Monte Carlo simulation that requires anything other than a uniform distribution, we have to convert the raw uniform variates to our target distribution, or ensure otherwise that we are meeting our distributional requirements. In this chapter, I discuss some of the known methods available for this purpose.
9.1
Inversion of the cumulative probability function
The cumulative probability function of any distribution has the following useful feature: For any variate x ∈ R from a given target distribution density ψ(x), the cumulative probability function for x, Rx i.e. Ψ(x) = −∞ ψ(x0 ) dx0 , is a uniform variate on the unit interval. This is because the cumulative probability function is by value just a probability measure, which is uniform by definition. So, if we can invert the cumulative probability function, and take values of the inverse cumulative probability from given uniform variates, we obtain variates of our target distribution! Example: The Cauchy distribution density and probability were given in equations (2.53) and (2.54) as ψ(x) =
1 1 π 1 + x2
Ψ(x) =
1 1 atan(x) + . π 2
The inverse cumulative density can easily be given: Ψ−1 (u) = tan π(u − 1/2 ) .
(9.1)
As for all distributions that are non-zero for all x ∈ R, the inverse cumulative probability function of the Cauchy distribution diverges both at u → 0 and at u → 1. It is therefore of paramount importance for numerical applications to ensure that the underlying uniform number generator never returns either 95
96
Non-uniform variates
0.35
1
y(x)
10
0.9
y(x)
0.3 0.8
5
0.25
0.7 0.6
0.2
0.5 0.15
Y -1(u)
0
0.4 0.3
0.1
-5 0.2
0.05 0.1 0 -10
-5
0
5
10
0 -10
-10 -5
0
5
10
0
0.2
0.4
0.6
0.8
1
Figure 9.1: Constructing Cauchy variates that are distributed according to ψ(x) as given in equation (2.53) from U(0, 1) variates is easy by virtue of its inverse cumulative probability Ψ−1 (u) in (9.1). of those two limiting values. Unfortunately, almost all uniform number generators that I have come across so far include at least 0 in their range, which must be intercepted in a Monte Carlo implementation. The method of direct inversion of the cumulative probability function is definitely the preferred method for non-uniform variate construction, wherever Ψ−1 (u) is readily available and can be computed efficiently. Despite the alternative methods discussed below, for distributions whose inverse cumulative probability function is not so easily computable, such as Student’s t whose cumulative is given in equation (2.52) or the GB2 distribution whose cumulative is (2.62), it may still be advisable to use the inverse cumulative mapping method. However, in those cases, rather than root-search for the inverse cumulative for each new draw, one should set up a (possibly multidimensional) interpolation table using cubic splines, monotonicity preserving splines, or your favourite shape-preserving spline method [PTVF92, Hym83, Kva00]. Fortunately, for most known distributions, we have at least the cumulative probability function and can thus readily set up the required interpolation table.
9.2 Using a sampler density There may be situations when the inverse cumulative probability function is either not available at all, or would be computationally extremely expensive to evaluate. In this case, we can employ the sampler density technique [Mac97]. Instead of drawing directly from the desired target density ψ(x), we choose another, hopefully similar, density ψe from which we can readily draw variates. The simulation is then carried out using variates from the sampler density, and each function evaluation is probability density e corrected according to the likelihood ratio of the target density ψ(x) and the sampler density ψ(x). Mathematically, this corresponds to the density transformation ! Z Z ψ(x) e f (x)ψ(x) dx = f (x) ψ(x) dx . e ψ(x)
(9.2)
97
9.2 Using a sampler density
In other words, the Monte Carlo estimator that was given in equation (2.15) is replaced by the sampler density Monte Carlo estimator vˆN
N 1 X f (xi ) := N i=1
ψ(xi ) e i) ψ(x
! .
(9.3)
Equation (9.3) gives us an immediate condition on any choice of sampler density: The sampler density e ψ(x) must not be zero wherever f (x)ψ(x) is non-zero! The attentive reader may have noticed that for any finite number N of simulations, the sampler density Monte Carlo estimator for a constant function f (x) := c will, generically, not result in the exact value c, unlike equation (2.15). This can be remedied by re-normalisation, which gives us the normalised sampler density Monte Carlo estimator: N P
vˆN :=
f (xi )
i=1
ψ(xi ) e i) ψ(x
(9.4)
N P ψ(xi ) i=1
e i) ψ(x
However, by virtue of the continuous mapping theorem, both estimators are valid. In general, the
y(x)
y~(x)
f(x)
x Figure 9.2: The sampler density method uses variates from a different, but preferably similar density, and corrects the average for the misrepresentation of the individual draws by multiplying each function evaluation by the likelihood ratio of the target and the sampler density. variance of any sampler density estimator will be different from a straight target density estimator, and by virtue of Murphy’s law, if we selected the sampler density more or less randomly, we will end up with an increased variance for our Monte Carlo estimator, i.e. a larger Monte Carlo error. What’s more, this problem is geometrically compounded as the number of dimensions increases, and this is the reason why drawing from the target distribution is to be preferred if it is readily available. For
98
Non-uniform variates
a more mathematical analysis of the problem of increased variance of the sampler density estimator, see [Mac97]. To finish this section, I show in figure 9.3 the convergence diagram for a standard plain vanilla European call option in the Black-Scholes framework with S0 = 123, K = 134, r = 12%, d = 4%, T = 3.33, and σ = 67% using the Cauchy density as the sampling distribution in comparison to the direct use of standard normal variates. The first thing to notice is that there is not much between the 55.4 Black-Scholes price Cauchy sampling density, not normalised Cauchy sampling density, normalised standard normal sampling density
55.2
running average
55
54.8
54.6
54.4
54.2 0
1000000
2000000
3000000 # of iterations
4000000
5000000
Figure 9.3: Pricing an out-of-the-money call option using a Cauchy sampling density. normalised and the un-normalised versions of the sampling density method in equations (9.4) and (9.3), respectively. There is one surprising feature though: the Cauchy sampler density method appears to converge faster for this particular simulation! The explanation is rather simple. In the Black-Scholes framework, the terminal value construction method for ST given S0 and a standard normal variate x is 1
ST = S0 e(r−d− 2 σ
2 )T +σ
√
Tx
.
Solving ST = K for x, we have xK =
ln(K/S0 ) − (r − d − 12 σ 2 )T √ . σ T
For the parameters given above, the cumulative normal probability for xK is N(xK ) = 67.85%. The cumulative Cauchy probability for xK is, however, only ΨCauchy (xK ) = 63.82%. This means that
99
9.2 Using a sampler density
approximately 4% fewer of the constructed ST values end up out of the money when we use the Cauchy sampling density, which gives rise to an improved convergence. What’s more, since the Cauchy density has much fatter tails than the normal distribution, the simulations that end up in the money tend to have higher values and thus compensate the convexity adjustment, i.e. the Itˆo term − 12 σ 2 T , more rapidly. Both effects together amount to the improved convergence behaviour seen in figure 9.3.
9.2.1
Importance sampling
Since we have a lot of liberty with respect to the choice of the sampler density, can we take advantage of this freedom in order to reduce the variance of the Monte Carlo estimator? Consider the special choice f (x)ψ(x) e , ψ(x) := R f (x)ψ(x) dx
(9.5)
ignoring for now the fact that we don’t actually know the value of the denominator. This particular choice for the sampler density would enable us to calculate (9.2) with any single draw for x since we obtain from equation (9.3) vˆ1 = f (x1 )
ψ(x1 ) e 1) ψ(x
!
R = f (x1 )ψ(x1 )
f (x)ψ(x) dx f (x1 )ψ(x1 )
Z =
f (x)ψ(x) dx .
Since we don’t know the solution of the problem that shows up in the denominator of the right hand side of equation (9.5), we clearly can’t actually do this. However, we can choose a sampler density that takes structural features of the product f (x)ψ(x) into account. A very simple choice that will already provide a significant improvement is to choose a sampler density that is zero wherever f (x) is zero, too. An example for an integrand f (x) that has regions of zero value is given with the piecewise affine function in figure 9.2, and in finance, functions with this feature are everywhere. Specific choices of sampler densities that take thus advantage of the regions of importance of the integrand f (x) are known as importance sampling methods. The use of a heavy-tailed distribution for the pricing of an out-of-the-money option as in the example in the previous section is one such application of importance sampling. We will revisit the importance sampling method in sections 10.5 and 11.4. Incidentally, the example of the ideal sampler density in equation (9.5) highlights that for strongly non-constant functions f , it may be advantageous to carry out the importance sampling technique using the non-normalised estimator given in equation (9.3).
100
Non-uniform variates
9.2.2
Rejection sampling
Rejection sampling is the stepsister of importance sampling. For this method, we need to select a e and a scaling constant c such that cψ(x) e sampling density ψ(x) ≥ ψ(x) for all x in the domain of ψ. In order to compute a Monte Carlo estimator over many values for x with x ∼ ψ, we proced as follows:e · Draw a variate x from the sampling density ψ(x). e · Compute the value of the sampling density ψ(x) at x, and also the value of the target density ψ(x) at x. · Draw a uniform variate u ∼ U(0, 1). e · If u·cψ(x) > ψ(x), reject this attempt to find a suitable variate and start again, otherwise accept x as a variate with x ∼ ψ and evaluate the integrand f (x) in the usual manner.
y
~ c (x)
y(x) y
~ u .c (x)
x Figure 9.4: For rejection sampling choose a sampling density ψe and establish a scaling constant c such e e that cψ(x) ≥ ψ(x) ∀ x. Then, take a draw x from the sampling density ψe and evaluate ψ(x) and ψ(x) e at x. Next, draw a uniform (0, 1) variate u. Use the drawn variate x if u·cψ(x) ≤ ψ(x), else reject it. There are two main problems with the rejection method. The first one is, again, the geometric implosion of the sampling yield as the number of dimensions increases1 . The second problem is that they are 1
This repeated occurrence of high numbers of dimensions causing problems with the evaluation of integrals gave rise to the term curse of dimensionality.
101
9.3 Normal variates
rather difficult to amend to the efficient application of low-discrepancy numbers. Personally, I don’t like rejection methods.
9.3
Normal variates
My favourite method for constructing standard normal variates is the highly sophisticated interpolation formula by Peter Acklam [Ack00] for the inverse cumulative normal distribution as discussed in section 2.3 on page 11. A very crude way to quickly construct (approximately) normally distributed variates is to add up 12 uniform variates, and subtract 6, and I have given a diagram that shows the difference to the exact normal distribution in figure 2.1 on page 20. For any reasonable application, I would always use either Peter Acklam’s method, or Boris Moro’s interpolation formula [Mor95].
9.3.1
The Box-Muller method
There is, however, yet another method for the construction of standard normal variates that is still very popular. It is based on the transformation (u, v) to (x, y) given by √ x = −2 ln u sin(2πv) √ . y = −2 ln u cos(2πv)
(9.6)
For u and v being independent standard uniform in (0, 1)2 , the joint distribution for x and y is given by ! ! − 12 y 2 − 12 x2 ∂(x, y) e e = √ √ ψ(x, y) = , (9.7) ∂(u, v) 2π 2π i.e. that of two independent standard normal variates! Techniques for the generation of standard normal variates based on the transformation (9.6) are known as Box-Muller methods [BM58, Knu81]. The classical application of the Box-Muller method is to draw two independent uniform variates, and transform them according to equation (9.6), in order to obtain two independent standard normal variates. Note that the trigonometric terms on the right hand side of equation (9.6) are the abscissa and ordinate of a point on the perimeter of a unit circle. Another variant of the Box-Muller method is to draw a random point from within a unit circle, and use its cartesian coordinates (s, t) as follows. First, set u := s2 + t2 . Then set q
x := s −2 lnuu q . y := t −2 lnuu
(9.8)
The advantage of this procedure is that no evaluation of trigonometric functions is required, and it used to be that those trigonometric functions were rather cpu-intensive in their evaluation (in comparison to simpler functions such as the logarithm and the square root)2 . The only question that remains is: how 2
More recently available computing hardware provides substantial improvements for a variety of previously cpu-time expensive functions, see section 14.6.
102
Non-uniform variates
do we draw a cartesian coordinate pair that describes a point inside a unit circle? The commonly used method for this purpose is a two-dimensional rejection procedure. We simply keep drawing uniform (−1, 1) variate pairs (by drawing standard uniform numbers, multiplying them by two, and subtracting 1) until we find one that lies inside the unit circle. Since the area of a unit circle is π, and the area of a 2 × 2 square is 4, the yield of this rejection method is π/4. There are two main problems with the Box-Muller method. The first one is discussed in the next section. The second problem is that rejection methods are highly dangereous (and should not be used, really) in conjunction with low-discrepancy numbers. An example for this is shown in figure 9.5. In the top-left diagram of the figure, two-dimensional Gaussian variates constructed from MersenneBox-Muller method by trigonometric transformation Box-Muller method by trigonometric transformation using Mersenne-Twister numbers. using Sobol’ numbers. 4 4
4
3
3
3
2
2
2
1
1
1
0
0
0
-1
-1
-1
-2
-2
-2
-3
-3
-3
-4 -4 4
-3
-2 -1 0 1 2 3 Inverse cumulative method using Mersenne-Twister numbers.
4
-4 -4 4
-3
-2 -1 0 1 2 3 Inverse cumulative method using 2-dimensional Sobol’ numbers.
4
-4 -4 4
3
3
3
2
2
2
1
1
1
0
0
0
-1
-1
-1
-2
-2
-2
-3
-3
-3
-4 -4
-3
-2
-1
0
1
2
3
4
-4 -4
-3
-2
-1
0
1
2
3
4
-4 -4
Box-Muller method by rejection using 2-dimensional Sobol’ numbers.
-3 -2 -1 0 1 2 3 Box-Muller method by rejection using 1-dimensional Sobol’ numbers.
-3
-2
-1
0
1
2
3
4
4
Figure 9.5: The Box-Muller method(s) in comparison to the use of the inverse cumulative normal function. All diagrams were constructed from 65535 pairs of variates. Twister numbers using equation (9.6) are shown. Next to it, the same transformation method was used with 2-dimensional Sobol’ numbers. Then, the distribution of 2-dimensional Sobol’ numbers transformed by the Box-Muller rejection method (9.8) are shown. At the bottom-left, we have the inverse cumulative normal function applied to Mersenne numbers, followed by the same method with 2-dimensional Sobol’ numbers. The last diagram highlights the danger of combining rejection methods with low-discrepancy numbers: using a one-dimensional sequential generator with the Box-Muller rejection method, as in this example, which is fine for pseudo-random numbers, goes horribly wrong when the number generator is a low-discrepancy algorithm.
103
9.3 Normal variates
9.3.2 The Neave effect A problem with highly sophisticated deterministic methods, such as pseudo-random number generators and the Box-Muller algorithm, is that it is often difficult to foresee when their interaction may have undesirable side-effects. In 1973, H. R. Neave discovered one of these hard-to-imagine quirks of undesirable interaction between nonlinear systems [Nea73]. When we use simple multiplicative congruential pseudo-random number generators such as Ran0 (also known as GGL3 ) as given in equation (7.6) and discussed in section 7.4 in conjunction with the transformation version of the Box-Muller method given in equation (9.6), there is a nasty surprise in store for us: over the whole period of the number generator of 231 − 1 = 2147483647 iterations, the smallest pseudo-normal variate that can be drawn for x in equation (9.6) is -4.476239, and the largest we will get is 4.717016 [Tez95]. Strictly speaking, there are two possible ranges for x depending on whether the pairwise transformation starts on the cycle containing the local seed 1, or on the cycle containing the local seed 16807 in the iteration mi = a · mi−1
mod M
with a = 16807, M = 231 − 1, and mi−1 being the local seed for the i-th single variate ui =
mi . M
However, the variation in the lower bound and the upper bound between those two possible cycles is beyond the first six decimal digits. According to the cumulative normal probability function, we would expect to have approximately 231 − 1 · N(−4.476239) ≈ 8157 draws lower than -4.476239, and equally several thousand above 4.717016. So, the range for x isn’t too good. The lower bound for y from equation (9.6), is -4.566210, again only showing variation between the two possible cycles beyond the first 6 decimal digits. The upper bound for y, though, is 6.385757568918 for the cycle beginning with mi = 1, and 6.555541555803 for the cycle beginning with 16807. In other words, the upper bound is somewhat better for the y variate, but the lower bound is still no good. What’s more, on either side of the distribution resulting from the combination of this number generator with the trigonometric version of the Box-Muller method, there appear to be some kind of wing formations in the tails of the distribution, as is shown in figure 9.6. As you can see, addition of a bitwise XOR mask as it is done for the proper Ran0 pseudo-random number method doesn’t fix the problem. The Neave effect does fade away, though, when we use the rejection version of the Box-Muller algorithm (9.8). However, there are reports that the rejection method also suffers from problems due to the fact that two (or more) uniform variates are used together for the generation of two normal variates [AW88, Rip87]. Of course, some may say that this is a small effect in the tails of the distribution, which is fair enough. However, in finance we are frequently particularly concerned with the tails since they tend to contain the most feared scenarios. But then, of course, you are probably not using the simple Ran0 aka 3
The only difference between the two is that Ran0 uses a bitwise XOR mask on the seed before and after each iteration mainly to prevent the accidental use of 0 as a seed, which would result in the fixed point 0.
104
Non-uniform variates
On a global scale, all combinations look pretty much the same 0.4 GGL : x GGL : y Ran0 (i.e. using a mask) : x Ran0 (i.e. using a mask) : y GGL with rejection method : x GGL with rejection method : y
0.35
0.00035
0.0003
0.3
0.00035
GGL : x GGL : y Ran0 (i.e. using a mask) : x Ran0 (i.e. using a mask) : y GGL with rejection method : x GGL with rejection method : y
GGL : x GGL : y Ran0 (i.e. using a mask) : x Ran0 (i.e. using a mask) : y GGL with rejection method : x GGL with rejection method : y
0.0003
0.00025
0.00025
0.0002
0.0002
0.00015
0.00015
0.0001
0.0001
5e-05
5e-05
0.25 0.2 0.15 0.1 0.05 0
-4
-2
0
2
4
0 -4.6
-4.5
-4.4
-4.3
-4.2
-4.1
-4
-3.9
-3.8
0 3.8
4
4.2
4.4
4.6
4.8
Figure 9.6: The Neave effect is due to an interaction of pseudo-random number generation methods and the trigonometric transformation version of the Box-Muller algorithm (9.6). In this figure, the GGL generator given by equation (7.6) with a = 16807, c = 0, and M = 231 − 1 was used. GGL generator, but one of the much more sophisticated ones that are readily available these days. Unfortunately, number theoreticians have begun to suspect that effects similar to the Neave phenomenon may occur for other number generation classes4 . The fundamental principle of the problem underlying the Neave effect remains, thus: whenever you use a variate-mixing transformation in order to generate variates from a specific target distribution, you run the risk of some kind of nonlinear interaction with the number generation mechanism that may be very hard to foresee, or even very difficult to notice, until something has gone wrong in a very big and unexpected way. In summary, since there are nowadays highly accurate and efficient interpolation algorithms available for the inverse cumulative normal probability function, it is generally safer to use those rather than to employ the Box-Muller method. It was an invention of great ingenuity and insight at the time, but now it has had its days.
9.4 Simulating multi-variate copula draws The problem of non-uniform variate generation is, of course, not limited to one dimension, or several dimensions of independent variates. The copula approach of creating multidimensional variate draws with co-dependence was explained in section 5.2. To conclude this chapter, I give in figure 9.7 an example for the kind of codependence that can be constructed by the use of copulæ. The Archimedean copula generating function φ(u) = (u−1 − 1)θ as given in equation (5.31) in section 5.2.3 was used with θ = 3/2. The Weibull variates were generated from the inverse cumulative Weibull probability function 1
Ψ−1 (u) = [−β ln(1 − u)] γ , 4
(9.9)
Tezuka states on page 152 in [Tez95] that the Neave effect ‘... possibly occurs not only with linear congruential sequences but also with LS(2) sequences.’
105
9.4 Simulating multi-variate copula draws 2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0 0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Figure 9.7: 131071 Weibull variates for β = 1 and γ = 4 under the Archimedean copula generated by φ given in equation (5.31) with θ = 3/2. which can be calculated from (2.72), using β = 1 and γ = 4. A rather interesting shape, don’t you think?
106
Non-uniform variates
Chapter 10 Variance reduction techniques As we have seen in chapter 2, the error estimate for the result of a Monte Carlo simulation is given by the standard deviation of the result (i.e. the average of all samples for a standard Monte Carlo integration) which is known as the standard error and usually estimated as the realised standard deviation of the simulation divided by the square root of the number of iterations. The smaller the variance (and thus the standard deviation) of the calculation, the more accurate will be the result. When I talk about variance here, I don’t mean the variance associated with any one individual sample of our Monte Carlo simulation, but with the variance associated with repeating the whole simulation many times. A number of techniques have been developed [BBG97], that help to reduce the variance of the result, and thus to reduce the number of simulations required for a given accuracy. Some of these methods can be combined to achieve even higher accuracy as we will see.
10.1 Antithetic sampling Whenever we use Gaussian variates to drive a Monte Carlo calculation, or wish to simulate Brownian motion by constructing sample paths of standard Wiener processes, we can make use of the fact that for any one drawn path its mirror image has equal probability. In other words, if a single evaluation driven by a Gaussian variate vector draw z i is given by vi = v(z i ), we also use v˜i = v(−z i ) . The standard error for antithetic sampling requires some consideration. This is because the central limit theorem underpinning the idea that the standard error can be estimated as the square root of the realised variance divided by the square root of the number of samples requires independent draws. However, if instead of viewing both vi = v(z i ) and v˜i = v(−z i ) as individual samples, we only count the pairwise average v¯i = 12 (v(z i ) + v(−z i )) as an individual sample, all is well again, because the pairwise averages v¯i are independent, and a standard error can be calculated from them in the ordinary fashion. Taking into account that each evaluation of v¯i requires two calculations (albeit only one Gaussian 107
108
Variance reduction techniques
vector draw generation), the antithetic sampling procedure provides a variance reduction if 1 1 V (vi + v˜i ) < V[vi ] 2 2
(10.1)
which is equivalent to Cov[vi , v˜i ] < 0 .
(10.2)
This is always the case if v(z) is monotonic in z. Whenever the first realised moment of the underlying variate draws {z i } has a strong impact on the result of the overall simulation, antithetic sampling is an easy way of improving the accuracy of the calculation since it corrects the first moment. Of course, if we use Monte Carlo simulation to calculate the value of derivatives contracts that mainly (or only) depend on higher moments, antithetic sampling will not help. Low-discrepancy numbers, or more specifically, Sobol’ numbers, unlike pseudo-random numbers, have the antithetic feature built into them, but only approximately. This is to say, whenever we use a recommended number of draws such as 2n − 1 for some n as advocated in section 8.7, the first moment of {z} is correct to within the numerical accuracy of the conversion from uniform (0, 1) to Gaussian variates. Also, it is worth remembering that low-discrepancy numbers are very carefully designed and tinkering with them can result in unexpected and rather undesirable effects. Therefore, even when we are not using 2n − 1 draws with low-discrepancy numbers, adding the antithetic method to the use of low-discrepancy numbers is unlikely to improve the accuracy, and instead can lead to erroneous results.
10.2 Variate recycling This method applies when the quantity to be calculated is essentially a re-evaluation of a functional of a (possibly multidimensional) discretised stochastic process, or any other re-evaluation using many draws. An example for this is the calculation of the Greeks of options by finite differencing. Let Wi be the discretised representation of our Monte Carlo approximation of a particular standard Wiener process path. Also, let v(p) = F [p, {Wi }i=1...N ] be the approximation of the value of an option by averaging the evaluation functional over the N sampling paths {Wi }i=1...N , depending on the parameter p. The simplest approach to estimate the sensitivity of v with respect to the parameter p is to run a separate Monte Carlo calculation using a slightly larger value for p, namely p + ∆p to obtain v(p + ∆p) and set ∂v ∂p
≈
v(p + ∆p) − v(p) . δp
(10.3)
In fact, for specific Greeks such as Vega, the dependence with respect to implied volatility, market practitioners may insist that it is mandatory to use forward differencing as in equation (10.3), and to use a particular increment ∆p (usually one absolute percentage point) for the calculation. Other Greeks where an absolute shift is commonly used are Rho, the sensitivity of FX and equity options with respect to domestic interest rates, and also the sensitivity with respect to forecast dividend yields.
109
10.3 Control variates
Naturally, we are interested in an accurate estimate of the Greeks when using the approach defined by equation (10.3). The standard measure for the accuracy of a Monte Carlo method is the variance of the resulting numerical approximaton. The variance of the sensitivity as given by equation (10.3) is thus
∂v V ∂p
≈
1 · E (v(p + ∆p) − v(p))2 − (E[(v(p + ∆p) − v(p))])2 ∆p
=
1 · E v(p + ∆p)2 − 2E[v(p + ∆p)v(p)] + E v(p)2 ∆p − (E[v(p + ∆p)])2 + 2E[v(p + ∆p)] E[v(p)] − (E[v(p)])2
= =
≈
(10.4)
1 · {V[v(p + ∆p)] + V[v(p)] − 2Cov[v(p + ∆p), v(p)]} ∆p 1 · {V[v(p + ∆p)] + V[v(p)] ∆p o p −2 V[v(p + ∆p)] V[v(p)]Corr[v(p + ∆p), v(p)] 2 · V[v(p)] {1 − Corr[v(p + ∆p), v(p)]} ∆p
(10.5)
where I used V[v(p + ∆p)] ≈ V[v(p)] in the last step. As equation (10.5) indicates, it is desirable to maximise the correlation of the two separate calculations v(p + ∆p) and v(p). For monotonic functions v(p), one can show [BBG97] that positive correlation is given if we use the same sample path set {Wi }i=1...N for both calculations whence one may call this method variate or path recycling. The above analysis easiliy transfers to other calculations of similar nature. Another example is the repeated Monte Carlo evaluation of the same problem with slightly varying parameters in the process of optimisation or within a nonlinear solver routine. Naturally, it is advisable to re-use (and wherever possible precalculate all quantities derived from them!) the sample paths, or simply the drawn variates if the problem doesn’t involve the concept of discretised stochastic processes.
10.3 Control variates Many Monte Carlo calculations are carried out for problems that we can almost solve analytically, or that are very similar to other problems for which we have closed form solutions. In this case, the use of control variates can be very beneficial indeed. The idea is as follows. Let’s assume that we wish to calculate the expectation E[v] of a function v(u) for some underlying vector draw u, and that there is a related function g(u) whose expectation g ∗ := E[g] we know exactly. Then, we have # " n n 1X 1X v(ui ) = E v(ui ) + β E n i=1 n i=1 "
!# n X 1 g∗ − g(ui ) n i=1
(10.6)
110
Variance reduction techniques
for any given β ∈ R and thus we can replace the ordinary Monte Carlo estimator n
1X v(ui ) n i=1
(10.7)
1X = [v(ui ) + β (g ∗ − g(ui ))] . n i=1
(10.8)
vˆ = by n
vˆCV The optimal choice of β is
β∗ =
Cov[v, g] V[g]
(10.9)
which minimises the variance of of vˆCV . Note that the function g(ui ) does not have to be the payoff of an analytically known option. It could also be the profit from a self-financing dynamic hedging strategy, i.e. a strategy that start with zero investment capital. For risk-neutral measures, the expected profit from any such strategy is zero which means that the control variate is simply the payoff from the dynamic hedging strategy along any one path. An intuitive understanding of the control variate method is to consider the case when v and g are positively correlated. For any draw v(ui ) that overestimates the result, g(ui ) is likely to overestimate g ∗ . As a result, the term multiplied by β in equation (10.8) is likely to correct the result by subtracting the aberration. The precise value of β ∗ is, of course, not known but can be estimated from the same simulation that is used to calculate vˆCV . As in all the situations when the parameters determining the result are calculated from the same simulation, this can introduce a bias that is difficult to estimate. In the limit of very large numbers of iterations, this bias vanishes, but the whole point of variance reduction techniques is to require fewer simulations and thus shorter run time. A remedy to the problem of bias due to a correlated estimate of the control parameter β is to use an initial simulation, possibly with fewer iterates than the main run, to estimate β ∗ in isolation. Fortunately, the control variate technique does usually provide such a substantial speedup in convergence that this initial parameter estimation simulation is affordable. However, for many applications, the magnitude of the bias is negligible. The easiest way to ascertain that there is no bias present that would be relevant for derivatives pricing purposes is to look at a convergence diagram of the simulation method, rather than a single result. The control variate method can be generalised to take advantage of more than one related closed form solution. However, this necessitates the estimation of more control parameters, and makes the method more susceptible to errors in their estimate. It is generally considered to be wiser to have one reasonable control variate than several mediocre ones. For instance, an option on a geometric average, which can be priced analytically for geometric Brownian motion, works exceedingly well as a control variate for arithmetic average options, whilst the use of both a standard European option and the underlying asset as joint control variates is only about as effective as the European option used as a control variate by itself [BG97b].
111
10.4 Stratified sampling
10.4
Stratified sampling
The idea here is to subdivide the sampling domain into smaller areas for each of which a representative value of the function is selected.
This can be particularly useful if a good approximation for the
Figure 10.1: An example for stratification with representative values for each segment.
average over small subdomains is available. Stratified sampling is conceptually akin to fixed lattice methods. It can also be of advantage when an assessment of the total probability of a small subdomain is difficult, and each evaluation in this domain is rather cpu time expensive, but it is known that the function which is being sampled varies very little in any one given subdomain. Whenever the probability associated with each segment of the stratification can be well approximated, stratified sampling can be used to evaluate the Monte Carlo integral by simpling calculating the weighted sum over the representative values, which makes it essentially some kind of (possibly irregular) lattice method. The segments into which the subdomain is partitioned don’t have to be of equal size. A better choice of stratification is to make the subdomains have approximately equal probability associated with them. However, the biggest problem with stratified sampling is that it is very difficult to obtain any kind of error estimate. Clearly, a statistical error estimate like the conventional standard error for pseudorandom sampling can be very misleading since the individual function values resulting from each draw are not independent variates in the usual meaning of the word. Also, the accuracy of any one calculation is limited by the stratification whence taking more and more samples will not make the result eventually converge to the exact answer as it would for a conventional Monte Carlo method following Kolmogorov’s strong law. In other words, the very technique that is supposed to increase convergence, i.e. the stratification itself, can introduce a finite bias of unknown sign.
112
Variance reduction techniques
10.5
Importance sampling
The concept of importance sampling is to focus on those regions which contribute most to the average of a Monte Carlo integration procedure. The most common use of this method in finance is to ensure that all drawn samples are in regions where the function to be evaluated is non-zero. This is the particular difficulty of out-of-the money option pricing. The standard procedure of generating paths would result in most evaluations resulting in zero payoff and are thus effectively a waste of cpu time. The main drawback of importance sampling is that it requires additional knowledge of the underlying problem. However, for very specific calculations, it can make a tremendeous difference in the convergence speed. Take for instance the function f (X, Y ) of two independent standard normal variates X and Y in figure 10.2.
A Monte Carlo integration of such a function will converge substantially
1.2 f(X,Y) 1 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0
0.8 0.6 0.4
Y
0.2 -0.2
0
0.2
0 0.4
0.6
0.8
1
-0.2 1.2
X Figure 10.2: The integration of localised functions can be accelerated by the use of importance sampling. faster if we restrict the normal variate draws to be in the subdomain where f is non-zero, in this case (X, Y ) ∈ [0, 1]2 . For the specific example here, this is easily done by the use of our knowledge of the probability p associated with the domain (X, Y ) ∈ [0, 1]2 , namely p = (N(1) − N(0))2 ,
113
10.6 Moment matching
by which the result of the conditioned Monte Carlo integration simply has to be multiplied. The construction of variates conditioned on the domain of importance is for the given example achieved by first drawing a vector of two uniform (0, 1) variates uX and uY , and constructing the importance sampling variates Xis and Yis according to Xis = N−1 [N(0) + (N(1) − N(0)) · uX ] and respectively for Yis . Another, more detailed, example for the potential of the importance sampling method is discussed in section 11.4.
10.6 Moment matching Moment matching used to be a very popular method before efficient and reliable low-discrepancy numbers became available. This method does usually give more accurate results for calculations that use pseudo-random numbers, although it is not guaranteed to do so. Assume a Monte Carlo simulation is to be carried out using a total of N variate vectors v of dimensionality d of a known joint distribution density ψ(v). Then, we can calculate the moments actually realised by the drawn variate vector set V := {vij } with i = 1..N and j = 1..d. The first moment for dimension j is given by N 1 X vij , hvij = N i=1
j = 1..d .
(10.10)
Using (10.10), we can construct a set of first-moment-corrected variates V˜ by subtraction of the average in each dimension, i.e. v˜ij = vij − hvij .
(10.11)
The realised covariance of the mean-corrected variate set can be concisely represented as C˜ = V˜ > V˜
(10.12)
if we view V˜ as a matrix whose rows comprise the individual d-dimensional mean-corrected vector draws. Using the same pseudo-square root decomposition approach as discussed in chapter 6, we can construct a new matrix Vˆ whose entries will meet the desired covariance C of the target distribution density ψ exactly. Define the elements of the desired covariance matrix C as Z cjk = vj vk ψ(v)dvj dvk .
(10.13)
Also, define the pseudo-square roots of both C and C˜ by C˜ = A˜ · A˜>
and
C = A · A> .
(10.14)
114
Variance reduction techniques
The correction matrix K that transforms V˜ to Vˆ can be computed by solving the linear system A˜> · K = A> ,
(10.15)
−1 K = A˜> · A> .
(10.16)
i.e.
It is easily possible for a covariance matrix not to be of full rank. This is the case whenever there are fewer driving factors than financial assets, for instance. This means that, whilst the original vector 0
0
draws v are d-dimensional, the target covariance matrix may be C ∈ Rd ×d with d0 > d, albeit that the rank of C is d. As a consequence, the matrix A will then have d0 rows, but only d columns, 0 0 i.e. A ∈ Rd ×d . However, A˜ will be of dimensions d × d, and K ∈ Rd×d . It is advisable to use a failsafe method for the solution of the linear system (10.15) such as the Moore-Penrose pseudo-inverse explained in section 2.9 in order to avoid problems due to the fact that there may be fewer dimensions in the vector v than there are variates in the target distribution. Another situation that may give rise to problems is when the number N of drawn vector variates is comparatively small, and the resulting realised covariance C˜ of the mean-corrected variate set is nearly singular. Then, the linear system (10.15) is ill-conditioned which can lead to rather unexpected results. In other words, I recommend −1 always to use the Moore-Penrose pseudo-inverse for A˜> in equation (10.16). Putting all of the above linear algebra together, we can convince ourselves that the covariancecorrected variate matrix Vˆ = V˜ · K
(10.17)
Vˆ > · Vˆ = C .
(10.18)
does indeed satisfy
Note that when using this method to correct the first and the second moment of a set of drawn variates it should be applied to the variates after having transformed them from the uniform (0, 1) distribution to whatever distribution is actually used, e.g. a joint normal distribution. This is because the nonlinearity in the transformation will have the effect that whilst you may have a set of uniform (0, 1) variates that have exactly the same mean and covariance matrix as in the continuous limit, the normal variates resulting from them after transformation will not meet the desired mean and covariance matrix of the joint normal distribution. As a simple example, take the very small set of only two onep p dimensional uniform (0, 1) variates {u1 , u2 } = { 1/2 − 1/12, 1/2 + 1/12}. The first and second moment of this set meet the moment of the continuous uniform (0, 1) distribution 1/2 and 1/3 exactly. Transforming (u1 , u2 ) to normal variates by the aid of the inverse cumulative normal distribution gives {z1 , z2 } = {N−1 (u1 ), N−1 (u2 )} = {−0.801832717, 0.801832717} which does have the desired first moment of exactly zero. However, the second moment of the transformed set is 0.64293571 which is very different from 1 as it should be for a standard normal variate.
115
10.6 Moment matching
The above example of an extremely small set of variates may appear somewhat construed. It does, however, highlight the dangers associated with moment matching. Correcting the first and second moment, or even higher moments for that matter, of a set of drawn variates that are to undergo further transformation, does not mean that the final distribution meets those moments equally. This cautioning note applies particularly to the common use of standard normal variates in the construction of geometric Brownian motion. The sheer fact that a discrete set of variates {z} has the same first and second moment as the continuous Gaussian distribution, does not mean that lognormal variates constructed according to ST = S0 eσˆ ·
√
T ·z
have a first and second moment in perfect agreement with the continuous
lognormal distribution. What’s more, even if we manage to correct the first few moments of the variate set representing the ultimate target distribution, the effects we inflict upon the higher moments are quite unpredictable. Most calculations are affected by more than just the first few moments: even comparatively simple problems such as the pricing of an option with a fairly vanilla payoff depend quite strongly on the third moment, for instance. These considerations bear particular relevance when we actually need to represent a distribution by a very small sample set such as in a stratification method, or in a tree. A discussion of the differences caused by first and second moment matching in the underlying normal variate space or the target lognormal variate space is given in section 13.7 and 13.6 in the context of non-recombining trees for the pricing of interest rate derivatives in the BGM/J framework. As for the standard error estimate when we use pseudo-random numbers in conjunction with moment-matching, this is somewhat problematic since the variates can no longer be considered independent. This directly affects another nice feature of Monte Carlo simulations: with ordinary Monte Carlo simulations, it is possible to continue drawing variates and monitor the running error estimate until the latter has dropped below a specified level of required accuracy. Not only is it no longer justified to use the running standard error as a statistical error measure since any new draw is no longer strictly independent (since we would have to re-match the moments), but since this would also involve re-computing the correction matrix K from equation (10.16) in each step, it would clearly become computationally prohibitively expensive. As a matter of fact, the one-off calculation of the realised covariance matrix (10.12) can easily be extremely cpu intensive since it grows like the square of the total dimensionality of the problem, and linearly in the number of vector draws (i.e. paths for option pricing) that are used. Should you ever use the moment matching method, you’d be well advised to monitor the cpu time that is actually being spent in total in the correction step, in comparison to the time spent in the main Monte Carlo simulation. For example, for multi-asset options of Asian style, the total dimensionality (which is the product of the number of time steps and the number of underlyings) can easily be moderately large. Since the computing effort grows like the square of d, it is not uncommon for such calculations to spend 90% or more of the total calculation time just in the moment matching part of the simulation, in particular since those multi-asset calculations frequently involve the evaluation of a very simple payoff function. In comparison to the use of low-discrepancy numbers, the moment matching method with pseudo-random numbers rarely provides a substantial speedup in convergence as a function of the number of simulations which means that the same accuracy could be
116
Variance reduction techniques
achieved by slightly increasing the number of iterations using, e.g., Sobol’ numbers. It is important to remember that any user’s perception of Monte Carlo simulation convergence is mainly a function of the time spent waiting for the result. Therefore, the moment matching method can easily make the Monte Carlo method appear slower for multi-asset, multi-time-stepped simulations. Finally, I ought to mention that moment matching should not be combined with the use of lowdiscrepancy methods. This is because their careful construction from number theoretical principles already tries to match all the moments in a well-balanced way and interfering with them can have unexpected effects. A simple way to ensure that the first moment is exactly met when using Sobol’ numbers, for instance, is to use N = (2n − 1) vector draws for some positive integer n, as explained in section 8.7. You will also find that the second moments are almost exactly met, especially when comparing to pseudo-random numbers. Since Sobol’ numbers are particularly fast to construct, I never found a situation when just using a few more draws with straightforward Sobol’ numbers did not outperform the use of moment matching with pseudo-random numbers, even for very low-dimensional problems. For high-dimensional, i.e. multi-asset simulation problems, importance-aware path construction as outlined in section 10.8 in order to achieve an effective dimensionality reduction ensures that Sobol’ numbers still give a convergence improvement over pseudo-random numbers when measured just in terms of the number of iterations required until a certain accuracy is met. As discussed above, though, when measured in terms of cpu time (which is a much better measure), moment matching becomes prohibitively cpu time expensive for high-dimensional problems. Just in case the message is still not clear: Use the moment matching method at your own peril.
10.7 Latin hypercube sampling Latin hypercube sampling isn’t actually a Monte Carlo method. Latin hypercube sampling is a way to crash cars. Seriously. This technique is used when probing the sampling space is (quite literally) extremely expensive. Basically, a Latin hypercube sampling scheme is the attempt to place sampling points in a multi-dimensional stratification with as little overlap in all one-dimensional projections as possible. Imagine that you wish to evaluate the effect of four control parameters on the safety of the driver of a car measured by the impact forces experienced in a frontal collision. The control parameters could be, for instance, the angle of the steering wheel column, the elasticity of the back rest of the drivers seat, the rigidity of the front of the vehicle’s chassis, and the amount of leg space in front of the driver’s seat. For each of those parameters, you have chosen seven possible settings, which represents a four-dimensional stratification. Clearly, it is desirable not having to crash 74 = 2401 cars to get an idea what the optimal combination of settings would be. The Latin hypercube scheme is a systematic method to sample the stratified layer in each control parameter (at least) once. An example for such an arrangement is shown in figure 10.3 in all of the possible two-dimensional projections of the four-dimensional domain. Incidentally, the points shown in figure 10.3 were taken as the first seven points of a four-dimensional Sobol’ sequence which highlights another advantage of that particular number generation method: Sobol’ numbers have the Latin hypercube property built-in.
117
dimension 2
dimension 1
dimension 4
dimension 1
dimension 4
dimension 3
dimension 1
dimension 4
dimension 3
dimension 2
10.8 Path construction
dimension 2
dimension 3
Figure 10.3: The Latin hypercube arrangement of sampling points.
10.8
Path construction
In many applications, we need to construct a simulated discretised path of a standard Wiener process over a set {ti }, i = 1..n, points in time. We can view the values wi := W (ti ) of the Wiener process at those points in time as a vector of random variates. Since we are talking about a standard Wiener process starting at W (0) = 0, the global expectation of all of the wi as averaged over many simulated paths must be zero. The elements of their covariance matrix C, however, are given by cij = Cov[W (ti ), W (tj )] = min(ti , tj ) .
(10.19)
Given a vector z of independent Gaussian variates, we can transform them into a vector w representing a single simulated Wiener process path according to w =A·z ,
(10.20)
A · A> = C
(10.21)
provided that the matrix A satisfies
with the elements cij of C given by (10.19). The decomposition of C as in equation (10.21) is not unique and for Monte Carlo simulations driven by pseudo-random numbers, it is also completely irrelevant which method is used. The specific path construction technique employed does not directly have
118
Variance reduction techniques
an impact on the variance of the result of any Monte Carlo simulation. However, as I will elaborate below, a fortunate choice of the path construction method can aid in the reduction of what is called the effective dimensionality which is the key to unleashing the full potential of Sobol’ numbers, and can lead to a significantly improved convergence behaviour when compared with ordinary pseudorandom numbers. It is because of the effect that the choice of the path construction method has on the convergence behaviour of simulations using Sobol’ numbers that I discuss them in this chapter.
10.8.1
Incremental
Probably the simplest way to construct a Wiener process path is the incremental path construction. It can be seen as a direct application of the Markov property of a Wiener process. The construction is carried out by simply adding a new increment that is scaled according to the time step: wi+1 = wi +
p ∆ti+1 · zi ,
with
zi ∼ N (0, 1) .
(10.22)
The construction matrix of the incremental method is given by the Cholesky decomposition of the covariance matrix: √
Aincremental
∆t1
0
0
0
···
0
√ √ ∆t1 ∆t 0 0 · · · 0 2 √ √ √ ∆t2 ∆t3 0 · · · 0 ∆t1 = .. .. .. .. .. . . . . · · · . . .. .. .. .. .. .. . . . . . √ √ √ √ ∆t1 ∆t2 ∆t3 · · · · · · ∆tn
(10.23)
As you can see from equation (10.22), the incremental construction is an extremely fast way to convert a vector of n Gaussian variates into a standard Wiener path over n steps in time. Since all of the square roots can be precalculated, all that is involved for each path is a total of n multiplications, and n − 1 additions. It is because of this speed advantage that incremental path construction is the method of choice when pseudo-random numbers are used.
10.8.2
Spectral
In section 6.2, I explained how any symmetric positive definite matrix can be split into its spectral pseudo-square root. The same method can be used here to compute a spectral path construction matrix Aspectral that satisfies (10.21). It is also possible to use an approximation for the spectral discrete path construction matrix. This approximation is given by the spectral decomposition of a continuous Wiener path over the time interval (0, T ] into an orthogonal Hilbert basis. In order to construct a fully continuous Wiener path over any time interval, we would need an infinite number of driving Gaussian variates. The Gaussian
119
10.8 Path construction variates are the coefficients {zk } of the decomposition of W (t) in r ∞ 2k − 1 π 2 X sin ωk t zk W (t) = with ωk = . T k=1 ωk 2 T
(10.24)
The approximation of the element akl of the spectral construction matrix Aspectral is to be calculated as follows. First, populate an initial stage approximation that is simply given by the continuous counterpart for the path from t0 = 0 to tn = T : r a0ik
=
2 sin ωk ti T ωk
(10.25)
Now, the most important part of the path construction is that the variance of wi = W (ti ) must be equal to ti , i.e. V[wi ] = ti .
(10.26)
The initial stage approximation A0 does not satisfy this. This deficit can be rectified by defining τi as the sum of the squares of all of the entries of the i-th row of A0 : τi :=
n X
2
(a0ik )
(10.27)
k=1
The elements a ˜ik of the approximation for the discrete spectral path construction matrix A˜ are then given by r a ˜ik =
sin ωk ti √ ti 0 a = ti r 2k−1 . τi ik n P sin ωl ti 2 l=1
(10.28)
2l−1
The approximate spectral split calculated in this way for a 12-step discretisation of a Wiener process path from t = 0 to t = 3 is shown below.
A˜spectral =
0.2142 0.4141 0.6070 0.7897 0.9591 1.1121 1.2462 1.3591 1.4489 1.5140 1.5535 1.5727
0.2094 0.3771 0.4885 0.5265 0.4852 0.3707 0.2004 0.0000 −0.2001 −0.3694 −0.4825 −0.5242
0.1998 0.3091 0.2931 0.1579 −0.0411 −0.2224 −0.3115 −0.2718 −0.1200 0.0811 0.2486 0.3145
0.1860 0.2208 0.0867 −0.1128 −0.2231 −0.1589 0.0293 0.1942 0.2070 0.0580 −0.1363 −0.2247
0.1685 0.1257 −0.0674 −0.1755 −0.0670 0.1236 0.1612 0.0000 −0.1610 −0.1231 0.0666 0.1747
0.1479 0.0376 −0.1332 −0.0718 0.1136 0.1011 −0.0869 −0.1236 0.0546 0.1376 −0.0186 −0.1430
0.1252 −0.0319 −0.1127 0.0607 0.0961 −0.0855 −0.0736 0.1045 0.0462 −0.1165 −0.0157 0.1210
0.1011 −0.0754 −0.0405 0.1053 −0.0402 −0.0741 0.0967 0.0000 −0.0966 0.0739 0.0400 −0.1048
0.0766 −0.0909 0.0357 0.0465 −0.0919 0.0654 0.0121 −0.0799 0.0852 −0.0239 −0.0561 0.0925
0.0526 −0.0813 0.0771 −0.0416 −0.0108 0.0585 −0.0820 0.0715 −0.0316 −0.0214 0.0654 −0.0828
0.0299 −0.0539 0.0698 −0.0752 0.0693 −0.0530 0.0286 0.0000 −0.0286 0.0528 −0.0689 0.0749
0.0093 −0.0180 0.0264 −0.0343 0.0417 −0.0484 0.0542 −0.0591 0.0630 −0.0658 0.0675 −0.0684
0.1369 −0.0172 −0.1348 0.0341 0.1305 −0.0505 −0.1241 0.0661 0.1158 −0.0806 −0.1057 0.0939
0.1176 −0.0727 −0.0727 0.1176 0.0000 −0.1176 0.0727 0.0727 −0.1176 0.0000 0.1176 −0.0727
0.0964 −0.1033 0.0143 0.0879 −0.1085 0.0284 0.0781 −0.1121 0.0420 0.0671 −0.1139 0.0550
0.0736 −0.1073 0.0829 −0.0135 −0.0632 0.1056 −0.0908 0.0267 0.0518 −0.1023 0.0973 −0.0396
0.0497 −0.0872 0.1030 −0.0934 0.0607 −0.0129 −0.0380 0.0796 −0.1014 0.0982 −0.0707 0.0257
0.0251 −0.0486 0.0690 −0.0851 0.0959 −0.1006 0.0990 −0.0912 0.0777 −0.0592 0.0371 −0.0126
(10.29)
This is to be compared with an exact spectral split:
Aspectral =
0.1996 0.3961 0.5863 0.7672 0.9361 1.0902 1.2271 1.3447 1.4410 1.5146 1.5644 1.5895
0.1965 0.3653 0.4829 0.5326 0.5076 0.4112 0.2571 0.0669 −0.1327 −0.3137 −0.4506 −0.5242
0.1902 0.3078 0.3078 0.1902 0.0000 −0.1902 −0.3078 −0.3078 −0.1902 0.0000 0.1902 0.3078
0.1810 0.2307 0.1131 −0.0865 −0.2234 −0.1983 −0.0294 0.1608 0.2344 0.1380 −0.0584 −0.2125
0.1689 0.1438 −0.0464 −0.1833 −0.1097 0.0899 0.1863 0.0687 −0.1278 −0.1775 −0.0234 0.1576
0.1541 0.0578 −0.1325 −0.1074 0.0922 0.1420 −0.0390 −0.1566 −0.0197 0.1492 0.0756 −0.1209
(10.30)
120
Variance reduction techniques
The effective covariance of the approximate spectral construction matrix (10.29) is:
A˜spectral · A˜> spectral =
0.2500 0.2726 0.2669 0.2668 0.2656 0.2655 0.2650 0.2648 0.2646 0.2645 0.2644 0.2654
0.2726 0.5000 0.5247 0.5187 0.5187 0.5172 0.5169 0.5162 0.5160 0.5156 0.5155 0.5173
0.2669 0.5247 0.7500 0.7757 0.7698 0.7699 0.7682 0.7680 0.7672 0.7670 0.7665 0.7695
0.2668 0.5187 0.7757 1.0000 1.0263 1.0205 1.0208 1.0191 1.0188 1.0180 1.0179 1.0214
0.2656 0.5187 0.7698 1.0263 1.2500 1.2766 1.2710 1.2714 1.2697 1.2696 1.2688 1.2738
0.2655 0.5172 0.7699 1.0205 1.2766 1.5000 1.5269 1.5214 1.5220 1.5203 1.5204 1.5255
0.2650 0.5169 0.7682 1.0208 1.2710 1.5269 1.7500 1.7771 1.7718 1.7725 1.7709 1.7782
0.2648 0.5162 0.7680 1.0191 1.2714 1.5214 1.7771 2.0000 2.0273 2.0220 2.0232 2.0294
0.2646 0.5160 0.7672 1.0188 1.2697 1.5220 1.7718 2.0273 2.2500 2.2777 2.2724 2.2832
0.2645 0.5156 0.7670 1.0180 1.2696 1.5203 1.7725 2.0220 2.2777 2.5000 2.5285 2.5323
0.2644 0.5155 0.7665 1.0179 1.2688 1.5204 1.7709 2.0232 2.2724 2.5285 2.7500 2.7928
0.2654 0.5173 0.7695 1.0214 1.2738 1.5255 1.7782 2.0294 2.2832 2.5323 2.7928 3.0000
(10.31)
As you can see, the diagonal elements meet the requirements exactly. The off-diagonal elements, however, indicate that a simulation based on this approximate spectral split would effectively simulate the realisations of the standard Wiener process to have a somewhat exaggerated correlation. Considering that we spend significant effort elsewhere to ensure that the realised variances and covariances of all our random variates meet the specifications given by any financial model as closely as possible, the approximate spectral path construction may seem a little bit too inaccurate. There is, however, another important lesson we can learn from the approximate spectral decomposition. The eigenvalues of the spectral decomposition are given by λk = and thus decay ∝ O
1 k2
1 , ωk2
(10.32)
. Since the eigenvalues of a discrete path covariance matrix are well approx-
imated by those of the continuous counterpart, we can conclude that these, too, decay very quickly. As a matter of fact, for a completely uniform spacing of monitoring times such as the one used above with ∆t = 1/4, both eigenvalues and eigenvectors can be derived analytically, which is shown in appendix 10.9.1. The importance of the eigenvalues is given by the fact that they directly represent the amount of variance that can be reproduced by only using a smaller subset of the orthogonal eigenvectors and thus to attempt to mimic an effectively n-dimensional variate draw (namely the standard Wiener process path over n points) by only using m Gaussian variates, with m < n. In statistics, there is the notion of the variability explained by just using the first m column vectors in a complete path construction matrix. It is given by the sum of all of the squares of the elements of the vectors used. For the spectral decomposition, this is just the sum of the eigenvalues as given in (10.32). In this sense, the spectral decomposition (given that the eigenvectors are sorted by decreasing size of their associated eigenvalues) is the optimal way to assign most importance to the first Gaussian variates in any given vector draw z. As it happens, this kind of effective dimensionality reduction is precisely what makes low-discrepancy numbers converge the fastest, so this might be the best path construction method of choice in conjunction with Sobol’ numbers. If only there wasn’t always some catch. In this case the problem is as follows. The spectral path construction method may provide the fastest convergence in conjunction with Sobol’ numbers as a function of the number of iterations that are carried out. However, more important than that is the amount of time spent in the simulation. Apart from the fact that the calculation of the (accurate) spectral path construction matrix is effectively a task involving
121
10.8 Path construction
O (n3 ) mathematical operations, the actual use of the matrix Aspectral during the simulation involves n2 multiplications and n(n − 1) additions for the construction of each and every path. When n is in the hundreds, this means that for a simulation involving possibly several tens of thousands of iterations we end up spending most of our time just in the transformation from standard normal variates to Wiener path coordinates. In the next section, I discuss my favourite path construction method which gives almost the same effective dimensionality reduction as the spectral method, but at the expense of only about 3n multiplications and 2n additions for each constructed path!
10.8.3 The Brownian bridge Similar to the spectral path construction method, the Brownian bridge is a way to construct a discretised Wiener process path by using the first Gaussian variates in a vector draw z to shape the overall features of the path, and then add more and more of the fine structure. The very first variate z1 is used to determine the realisation of the Wiener path at the final time tn of our n-point discretisation of the path √ by setting Wtn = tn z1 . The next variate is then used to determine the value of the Wiener process as it was realised at an intermediate timestep tj conditional on the realisation at tn (and at t0 = 0 which is, of course, zero). The procedure is then repeated to gradually fill in all of the realisations of the Wiener process at all intermediate points, in an ever refining algorithm. In each step of the refinement procedure to determine Wtj given that we have already established Wti and Wtk with ti < tj < tk , we make use of the fact that the conditional distribution of Wtj is Gaussian with mean tj − ti tk − tj Wti + Wtk E Wtj = tk − ti tk − ti
(10.33)
and variance (tj − ti ) (tk − tj ) V Wtj = . (tk − ti )
(10.34)
A proof of equations (10.33) and (10.34) is sketched in appendix 10.9.2. Since all of the weighting coefficients can be precalculated, we only need to carry out three multiplications and two additions for each point in the Brownian bridge. Exceptions are, of course, the terminal point for which only a multiplication is required, and all those whose left hand side conditioning point is the beginning of the path at zero. The Brownian bridge is particularly easy to construct if the number of steps is a power of two because then each interval divides into two intervals containing the same number of steps. In general, we can construct the Brownian bridge by always subdividing the interval from the last point that was set to the next one that is already set, and halve this interval into two parts containing approximately the same number of points. An example for this procedure is illustrated in figure 10.4. The algorithm for the construction of a Brownian bridge over an arbitrary number of steps can be separated into two parts. The first part is the calculation of the indices determining the order of the point construction, the weighting coefficients, and the conditional variances, or their square roots, respectively. The second
122
Variance reduction techniques
9 5
1
2
4 6
3
8 7
10
0
1
2
3
12
11
4
13
5
6
7
8
9
10
14
11
12
13
14
Figure 10.4: The construction of a Brownian bridge over 14 steps. part is the mapping of a vector of standard Gaussian variates to a single Wiener path at run time of the Monte Carlo simulation. I give in code example 10.1 a concrete implementation of a C++ class providing these two stages. The constructor of the class carries out all of the initial calculations. The buildPath method is then to be called in the main loop of the Monte Carlo simulation. Each time the procedure buildPath is executed, it transforms a vector of uncorrelated standard normal variates given by the input vector normalVariates into a standard Wiener path with equal time steps of size ∆t = 1 and stores it in the output vector path. When we wish to construct paths that represent Brownian motion over points in time ti for i = 1..m, and the realisation of the Brownian path at each time horizon ti is to be consistent with an implied volatility coefficient σ ˆi , we need to take into account the resulting term structure of instantaneous volatility and variance. Given that we have a Brownian process B that is generated from a standard Wiener process W by the multiplication with a deterministic volatility function σ(t) according to dB = σ(t) dW ,
(10.35)
the conditional mean and variance for the realisation of B(tj ) given that we have knowledge of B(ti ) and B(tk ) with ti < tj < tk is E Btj =
vk − vj vk − vi
Bti +
vj − vi vk − vi
Btk
(10.36)
and (vj − vi ) (vk − vj ) V Btj = (vk − vi )
(10.37)
with Ztq vq = s=0
σ(s)2 ds .
(10.38)
10.8 Path construction
123
#include #include class BrownianBridge { // Builds Wiener process paths of equal time steps of delta t = 1. public: BrownianBridge( unsigned long numberOfSteps ); void buildPath( vector <double> &theWienerProcessPath, const vector <double> &gaussianVariates ); private: unsigned long numberOfSteps; vector leftIndex, rightIndex, bridgeIndex; vector <double> leftWeight, rightWeight, stddev; }; BrownianBridge::BrownianBridge( unsigned long numberOfSteps_ ) : numberOfSteps(numberOfSteps_), leftIndex(numberOfSteps), rightIndex(numberOfSteps), bridgeIndex(numberOfSteps), leftWeight(numberOfSteps), rightWeight(numberOfSteps), stddev(numberOfSteps) { assert(numberOfSteps); // There must be at least one step. vector map(numberOfSteps); // map is used to indicate which points are already constructed. If map[i] is zero, path point i // is yet unconstructed. map[i]-1 is the index of the variate that constructs the path point # i. unsigned long i,j,k,l; map[numberOfSteps-1] = 1; // The first point in the construction is the global step. bridgeIndex[0] = numberOfSteps-1; // The global step is constructed from the first variate. stddev[0] = sqrt(numberOfSteps); // The variance of the global step is numberOfSteps*1.0. leftWeight[0] = rightWeight[0] = 0.; // The global step to the last point in time is special. for (j=0,i=1;i>1); // l is now the index of the point to be constructed next. map[l]=i; bridgeIndex[i] = l; // The i-th Gaussian variate will be used to set point l. leftIndex[i] = j; // Point j-1 is the left strut of the bridge for point l. rightIndex[i] = k; // Point k is the right strut of the bridge for point l. leftWeight[i] = (k-l)/(k+1.-j); rightWeight[i] = (l+1.-j)/(k+1.-j); stddev[i] = sqrt(((l+1.-j)*(k-l))/(k+1.-j)); j=k+1; if (j>=numberOfSteps) j=0; // Wrap around. } } void BrownianBridge::buildPath( vector <double> &path, const vector <double> &normalVariates ){ assert( normalVariates.size() == numberOfSteps && path.size() == numberOfSteps ); unsigned long i,j,k,l; path[numberOfSteps-1] = stddev[0]*normalVariates[0]; // The global step. for (i=1;i
.
(12.19)
Using ∂Pi+1 τk = −Pi+1 · 1{k≥i} , ∂fk 1 + fk τk
(12.20)
where 1{k≥i} is one if k ≥ i and zero otherwise, and equations (12.15), (12.16), and (12.14), we have ∂SRi Pk+1 τk Nk τk Ak τk Ai Bk = − · + · · 1{k≥i} . (12.21) ∂fk Bi 1 + fk τk Bi 1 + fk τk Bi2
160
Monte Carlo in the BGM/J framework
This enables us to calculate the elements of the forward rate to swap rate covariance transformation matrix Z FRA→SR to obtain the expression FRA→SR Zik =
Pk+1 Nk fk τk Ai | {z }
constant weights approximation
+
(Ai Bk − Ak Bi )fk τk A B (1 + f τ ) | i i {z k k }
· 1{k≥i} .
(12.22)
shape correction
The second term inside the square brackets of equation (12.22) is called the shape correction. Rewriting it as n−1 k−1 X X (Ai Bk − Ak Bi )fk τk fk τk = · Pl+1 Pm+1 Nl Nm τl τm (fl − fm ) Ai Bi (1 + fk τk ) Ai Bi (1 + fk τk ) l=i m=k
(12.23)
highlights that it is a weighted average over inhomogeneities of the yield curve. In fact, for a flat yield curve, all of the terms (fl − fm ) are identically zero and the mapping matrix Z FRA→SR is equivalent to the constant-weights approximation (11.7) in [Reb99]. As things stand at this point, we have a map between the instantaneous FRA/FRA covariance matrix and the instantaneous swap/swap covariance matrix. Unfortunately, though, the map involves the state of the yield curve at any one given point in time via the matrix Z. The price of a European swaption, however, does not just depend on one single realised state or even path of instantaneous volatility. It is much more appropriate to think about some kind of path integral average volatility. Using arguments of factor decomposition and equal probability of up and down moves (in log space), Rebonato shows in [JR00] that the specific structure of the map allows us to approximate the effective implied swaption volatilities by simply using today’s state of the yield curve for the calculation of the mapping matrix Z: v RT u n−1 u X σk (t0 )σl (t0 )ρkl dt0 FRA→SR t σ ˆSRi (t, T ) = · ZilFRA→SR (0) Zik (0) · t T k=i,l=i
(12.24)
This approximate equivalent implied volatility can now be substituted into the Black swaption formula to produce a price without the need for a single simulation! In practice, the formula (12.24) works remarkably well. This is demonstrated for a whole sequence of coterminal European swaptions out in figure 12.4. An explanation for the remarkable accuracy of approximation (12.24) is beyond the scope of this section but can be found in [JR00]. Using the above preliminaries, I now outline the calibration procedure in detail. For a given time step from t to T , populate a time-unscaled FRA/FRA covariance matrix RT σk (t0 )σl (t0 )%kl (t0 )dt0 FRA t Ckl = . (T − t)
(12.25)
Next, map this matrix into a time-unscaled swap/swap covariance matrix using the Z matrix calculated from the initial state of the yield curve C SR = Z · C FRA · Z >
(12.26)
161
12.4 Calibration to European swaptions 45 678
/0 1(23
Y[ZM\^]
*+ ,-.
$% &(')
_a`bdcfeaghjifk lnm^oqprsutfv wyx{z|~}f
!"#
^jffR R^
adfajf
n¡f¢(£j¤n¥{¦ §~¨
q ©ª
«^¬jf®f¯R°±² ³R´^µ
9
:
;
@?
ACB
DCE
FHG
IHJ
KML
NPORQ STVU W X
Figure 12.4: The pricing error from equation (12.24) using only the constant weights approximation in the formula (12.22) or when including the shape correction for a GBP yield curve for August 10th , 2000. and calculate from it the swap/swap correlation matrix RSR given by CijSR
SR Rij =q
.
(12.27)
SR CiiSR Cjj
Now, we compute the spectral pseudo square root B of RSR which satisifies RSR = B SR B SR
>
.
(12.28)
At this point, we take into account the market given swaption prices. Denote the market implied market volatility of the swaption expiring at time th by σSR and define the diagonal matrix Ξ by h market Ξgh = σ ˆSR · h
σ ˆSRh (t, T ) · δgh σ ˆSRh (0, th )
(12.29)
with δgh being the Kronecker symbol (which is zero unless g = h when it is one) and both σ ˆSRh (t, T ) and σ ˆSRh (0, th ) calculated from the FRA instantaneous volatility parametrisation through equation (12.24). The final step is now to construct the FRA driver4 matrix AFRA by scaling up the swap/swap correlation driver matrix B SR and mapping back to FRA coordinates: AFRA = Z −1 · Ξ · B SR . 4
Karatzas and Shreve call this matrix the dispersion matrix [KS91] (page 284).
(12.30)
162
Monte Carlo in the BGM/J framework
The effective (time-unscaled) FRA/FRA covariance matrix is finally FRA Ceffective = AFRA AFRA
>
.
(12.31)
FRA In order to use the matrices Ceffective and AFRA for the evolution over the time step t → T from a set p of standard normal variates, we still need to multiply them by (T − t) and (T − t), respectively.
Within the limits of the approximation (12.24), using these matrices wherever we have expressions involving aij and cii in the following sections, will provide calibration to European swaption prices whilst retaining as much calibration to the caplets as is possible without violating the overall FRA/FRA correlation structure too much which is typically exactly what a practitioner wants for the pricing of Bermudan swaptions. As a side note, it should be mentioned that various other combinations of FRA carrying out the split into Ceffective and AFRA are possible, whilst still preserving calibration to European
swaptions. In my experience, however, the above approach represents the best method to preserve as much of the overall FRA/FRA correlation structure in the calibration as possible.
12.5 The Predictor-Corrector scheme In order to price a Bermudan swaption in a Monte Carlo framework, we need to evolve the set of forward rates f from its present values into the future according to the stochastic differential equation dfi (t) = fi (t) · µi (f (t), t)dt + fi (t) ·
m X
aij dWj
(12.32)
j=1
driven by an m-dimensional standard Wiener process W . The drift terms given by equation (12.8) are clearly state-dependent and thus indirectly stochastic which forces us to use a numerical scheme to solve equation (12.32) along any one path. Ideally, we would want to evolve the forward yield curve f only over the points in time which we actually need to monitor, i.e. the possible exercise dates. The simplest numerical scheme for the integration of stochastic differential equations5 is the Euler method fiEuler (f (t), t
+ ∆t) = fi (t) + fi (t) · µi (f (t), t)∆t + fi (t) ·
m X
√ aij (t)zj ∆t
(12.33)
j=1
with zj being m independent normal variates. This would imply that we approximate the drift as constant over the time step t → t + ∆t. Moreover, this scheme effectively means that we are using a normal distribution for the evolution of the forward rates over this time step. Whilst we may agree to the approximation of a piecewise constant (in time) drift coefficient µi , the normal distribution may be undesirable, especially if we envisage to use large time steps ∆t for reasons of computational efficiency. However, when we assume piecewise constant drift, we might as well carry out the integration over the time step ∆t analytically and use the scheme fiConstant drift (f (t), t 5
µi (f (t),t)∆t− 21 cii +
+ ∆t) = fi (t) · e
m P
aij zj
j=1
See [KP99] for a whole variety of methods for the integration of stochastic differential equations.
(12.34)
163
12.5 The Predictor-Corrector scheme √
∆t for A and by ∆t for C has been absorbed into the respective √ matrices. In other words, I have set A0 := A · ∆t and C 0 := C · ∆t and dropped the primes. Equation whereby the time step scaling by
(12.34) can also be viewed as the Euler scheme in logarithmic coordinates. The above procedure works very well as long as the time steps ∆t are not too long and is widely used and also referred to in publications [And00, GZ99]. Since the drift term appearing in the exponential function in equation (12.34) is in some sense a stochastic quantity itself, we will begin to notice that we are ignoring Jensen’s inequality when the term µi ∆t becomes large enough. This happens when we choose a big step ∆t, or the forward rates themselves or their volatility are large. Therefore, we should use a predictor-corrector method which models the drift as indirectly stochastic [HJJ01]. In the notation of Kloeden and Platen [KP99], depending on the time dependence of the instantaneous volatility function, this is an explicit order 2.0 weak scheme or order 1.0 weak predictor-corrector method6 . This combination of the Euler scheme in logarithmic space with the predictor-corrector method for the drift results in a remarkable accuracy and is used throughout all of the calculations reported in this chapter. The method is as follows. First, given a current evolution of the yield curve denoted by f (t), we calculate the predicted solution f Constant drift (f (t), t + ∆t) using one m-dimensional normal variate draw z following equation (12.34). Next, we recalculate the drift using this evolved yield curve. The predictor-corrector approximation µ ˜i for the drift is then given by the average of these two calculated drifts, i.e. µ ˜i (f (t), t → t + ∆t) = 1/2 µi (f (t), t) + µi (f Constant drift (f (t), t + ∆t), t) .
(12.35)
Finally, the predictor-corrector evolution is given by fiPredictor-corrector (f (t), t
µ ˜i (f (t),t → t+∆t)∆t− 12 cii +
+ ∆t) = fi (t) · e
m P j=1
aij zj
(12.36)
wherein we re-use the same normal variate draw z, i.e. we only correct the drift of the predicted solution. A hand-waving reasoning for the above approximation goes as follows. If we had to choose, for the calculation of a constant drift approximation, for any one time step, whether we use the initial forward rates f (t), or those at the end of the time step f (t + ∆t), neither of them appears to be superior over the other for the job. This is equivalent to the considerations about explicit and implicit methods for the numerical solution of both ordinary and partial differential equations. We don’t actually know the drift at the end of a desired time step, and solving for it as we would in an implicit method would 6
To be precise, a hybrid method is used here. In the approach presented here, I integrate in equation (12.25) the volatility functions indepently over the time step to obtain an equivalent discrete time step covariance matrix, and then treat these covariance matrices as if they had been given by a process of volatility functions that are constant in time. Therefore, neither is the explicit order 2.0 weak scheme given by equation (15.1.4) in [KP99] used, nor the order 1.0 weak predictor-corrector method as in equation (15.5.4) [KP99]. In the case of constant, i.e. time-independent volatility, however, these two schemes are identical. Thus, following the general notion that it is always beneficial to use as many explicit analytical solutions as possible in any numerical scheme, one can say that we are using a predictor-corrector scheme only for the drift term, not for the entire stochastic differential equation driving the evolution.
164
Monte Carlo in the BGM/J framework
require solving a high-dimensional non-linear system each and every time. However, we can approximate the drift term at the end of the time step, and then take the average of the two individual drift approximations, in analogy to the predictor-corrector method used in other areas of numerical analysis. A number of numerical experiments confirmed that this method is very robust and works very well (the error is never more than a fraction of the bid-offer spread which is typically around one vega of the option) for the prevailing interest rates and volatility levels for all of the major markets, even for very long dated (tens of years) options when only a single step to maturity is used. In order to demonstrate this, it is shown in figures 12.5 and 12.6 how the predictor-corrector drift approach performs for a Libor-in-arrears scenario in comparison to the piecewise constant drift approach where the drift term over any one time step is given by the state of the yield curve at the beginning of the step. In both figures, the error in the expectation of the Libor-in-arrears contract for both stepping methods
"!$#% '& ()! * ,+"-
dfe?gihkjl%gijm?npo qrj s n tum?o vQjeunpwpv6eunpnpt%v6j3e?n xfy{z v6l%|?} t%j~t%il
τ
/$03:6/$5
/$0 9"/$5 c
aX Y \UW
/$0386/$5 b
[T Xa U `\ ^ _]
/$0376/$5
[\ UZ ZYZ
XY V WU
/$0326/$5
RS T /$0 /"/$5
./10324/15
;
discounted expectation 25%
Parametrised exercise boundary for 32768 training paths Parametrised exercise boundary for 131072 training paths Parametrised exercise boundary for 1048576 training paths
s1
20%
15%
10%
5%
0% 0%
5%
10%
15% f0
20%
25%
30%
Figure 12.11: The exercise domain in the fi -si+1 projection of the evolved yield curve at ti = 2, together with the parametrised exercise boundary resulting from training with different sizes of the training set. only weakly depends on slight changes to the boundary. After all, if we view the price approximation as a function of the location of a given exercise boundary, then the real Bermudan option price is the one resulting from an optimisation over all possible exericse boundary locations. As it happens with the value of a function at its maximum, the first derivative with respect to its argument must be zero at the extremum, and thus the price approximation depends only weakly (i.e. as of second order) on minor changes of the exercise boundary location near the optimal point. Now, the results for a 6-non-call-2 semi-annual payer’s Bermudan swaption for a typical Sterling yield curve and volatility environment are presented. This means the yield curve was slightly downward sloping, and the implied swaption volatilities increased from 19.53% for the first one up to 22.46% for the last one (which is a caplet). The forward rates were again assumed to have piecewise constant instantaneous volatility but calibration was in this case done such that the entailed European swaptions’ prices are to be independent on the number of used factors. The option was again at the
171
12.9 Numerical results
money11 with a strike of 6.63%. In figure 12.12, a diagram of the Bermudan swaption price calculated using the presented Monte Carlo method in comparison to the values obtained from a non-recombining tree model is shown. For reference, the prices obtained from both models for the most expensive
6"798;:=?9@A BDCE@)F 8G7"HICJ:LK MN@M@)O P 87"7
5
243 01
6"798;:=?9@A QRCJ@ P 7LSD?98UT C
V-
